Professional Documents
Culture Documents
Eckhard Meinrenken
1
6
7
10
14
15
18
20
25
27
31
35
42
51
57
65
69
70
73
74
77
80
Remark 1.2. Using the implicit function theorem, one can show that smoothness of Inv is in
fact automatic. (Exercise)
The first example of a Lie group is the general linear group
GL(n, R) = {A Matn (R)| det(A) 6= 0}
of invertible n n matrices. It is an open subset of Matn (R), hence a submanifold, and the
smoothness of group multiplication follows since the product map for Matn (R) is obviously
smooth.
Our next example is the orthogonal group
O(n) = {A Matn (R)| AT A = I}.
To see that it is a Lie group, it suffices to show that O(n) is an embedded submanifold of
Matn (R). In order to construct submanifold charts, we use the exponential map of matrices
X
1 n
exp : Matn (R) Matn (R), B
7 exp(B) =
B
n!
n=0
d
|t=0 exp(tB) = B, hence the differential of exp
(an absolutely convergent series). One has dt
at 0 is the identity idMatn (R) . By the inverse function theorem, this means that there is > 0
such that exp restricts to a diffeomorphism from the open neighborhood U = {B : ||B|| < }
of 0 onto an open neighborhood exp(U ) of I. Let
is an embedded submanifold of GL(n, R), and hence is a Lie group. The submanifold charts
are obtained by exponentiating the subspace
sl(n, R) = {B Matn (R)| tr(B) = 0},
using the identity det(exp(B)) = exp(tr(B)).
Actually, we could have saved most of this work with O(n), SL(n, R) once we have the
following beautiful result of E. Cartan:
Fact: Every closed subgroup of a Lie group is an embedded submanifold, hence
is again a Lie group.
We will prove this very soon, once we have developed some more basics of Lie group theory.
A closed subgroup of GL(n, R) (for suitable n) is called a matrix Lie group. Let us now give a
few more examples of Lie groups, without detailed justifications.
Examples 1.3.
(a) Any finite-dimensional vector space V over R is a Lie group, with product
Mult given by addition.
(b) Let A be a finite-dimensional associative algebra over R, with unit 1A . Then the group
A of invertible elements is a Lie group. More generally, if n N we can create the
algebra Matn (A) of matrices with entries in A, and the general linear group
GL(n, A) := Matn (A)
is a Lie group. If A is commutative, one has a determinant map det : Matn (A) A,
and GL(n, A) is the pre-image of A . One may then define a special linear group
SL(n, A) = {g GL(n, A)| det(g) = 1}.
(c) We mostly have in mind the cases A = R, C, H. Here H is the algebra of quaternions
(due to Hamilton). Recall that H = R4 as a vector space, with elements (a, b, c, d) R4
written as
x = a + ib + jc + kd
with imaginary units i, j, k. The algebra structure is determined by
i2 = j 2 = k2 = 1, ij = k, jk = i, ki = j.
Note that H is non-commutative (e.g. ji = ij), hence SL(n, H) is not defined. On the
other hand, one can define complex conjugates
x = a ib jc kd
and
|x|2 := xx = a2 + b2 + c2 + d2 .
defines a norm x 7 |x|, with |x1 x2 | = |x1 ||x2 | just as for complex or real numbers. The
spaces Rn , Cn , Hn inherit norms, by putting
n
X
|xi |2 , x = (x1 , . . . , xn ).
||x||2 =
i=1
The subgroups of GL(n, R), GL(n, C), GL(n, H) preserving this norm (in the sense that
||Ax|| = ||x|| for all x) are denoted
O(n), U(n), Sp(n)
and are called the orthogonal, unitary, and symplectic group, respectively. Since the
norms of C, H coincide with those of C
= R4 , we have
= R2 , H = C2
U(n) = GL(n, C) O(2n), Sp(n) = GL(n, H) O(4n).
In particular, all of these groups are compact. One can also define
SO(n) = O(n) SL(n, R), SU(n) = U(n) SL(n, C),
these are called the special orthogonal and special unitary groups. The groups SO(n), SU(n), Sp(n)
are often called the classical groups (but this term is used a bit loosely).
e is again a Lie group. The universal cover
(d) For any Lie group G, its universal cover G
^R) is an example of a Lie group that is not isomorphic to a matrix Lie group.
SL(2,
1.2. Lie algebras.
Definition 1.4. A Lie algebra is a vector space g, together with a bilinear map [, ] : g g g
satisfying anti-symmetry
[, ] = [, ] for all , g,
and the Jacobi identity,
[, [, ]] + [, [, ]] + [, [, ]] = 0 for all , , g.
The map [, ] is called the Lie bracket. A morphism of Lie algebras g1 , g2 is a linear map
: g1 g2 preserving brackets.
The space
gl(n, R) = Matn (R)
is a Lie algebra, with bracket the commutator of matrices. (The notation indicates that we
think of Matn (R) as a Lie algebra, not as an algebra.)
A Lie subalgebra of gl(n, R), i.e. a subspace preserved under commutators, is called a matrix
Lie algebra. For instance,
sl(n, R) = {B Matn (R) : tr(B) = 0}
and
o(n) = {B Matn (R) : B T = B}
are matrix Lie algebras (as one easily verifies). It turns out that every finite-dimensional real
Lie algebra is isomorphic to a matrix Lie algebra (Ados theorem), but the proof is not easy.
The following examples of finite-dimensional Lie algebras correspond to our examples for Lie
groups. The origin of this correspondence will soon become clear.
Examples 1.5.
(a) Any vector space V is a Lie algebra for the zero bracket.
(b) Any associative algebra A can be viewed as a Lie algebra under commutator. Replacing
A with matrix algebras over A, it follows that gl(n, A) = Matn (A), is a Lie algebra, with
bracket the commutator. If A is commutative, then the subspace sl(n, A) gl(n, A) of
matrices of trace 0 is a Lie subalgebra.
i=1
and define o(n), u(n), sp(n) as the matrices in gl(n, R), gl(n, C), gl(n, H) satisfying
hBx, yi = hx, Byi
for all x, y. These are all Lie algebras called the (infinitesimal) orthogonal, unitary,
and symplectic Lie algebras. For R, C one can impose the additional condition tr(B) =
0, thus defining the special orthogonal and special unitary Lie algebras so(n), su(n).
Actually,
so(n) = o(n)
T
since B = B already implies tr(B) = 0.
Exercise 1.6. Show that Sp(n) can be characterized as follows. Let J U (2n) be the unitary
matrix
0
In
In 0
where In is the n n identity matrix. Then
Sp(n) = {A U(2n)| A = JAJ 1 }.
Here A is the componentwise complex conjugate of A.
Exercise 1.7. Let R() denote the 2 2 rotation matrix
cos sin
R() =
.
sin cos
Show that for all A SO(2m) there exists O SO(2m) such that OAO1 is of the block
diagonal form
R(1 )
0
0
0
0
R(2 ) 0
0
0
0
0 R(m )
For A SO(2m + 1) one has a similar block diagonal presentation, with m 2 2 blocks R(i )
and an extra 1 in the lower right corner. Conclude that SO(n) is connected.
Exercise 1.8. Let G be a connected Lie group, and U an open neighborhood of the group unit
e. Show that any g G can be written as a product g = g1 gN of elements gi U .
Exercise 1.9. Let : G H be a morphism of connected Lie groups, and assume that the
differential de : Te G Te H is bijective (resp. surjective). Show that is a covering (resp. surjective). Hint: Use Exercise 1.8.
(t)
T(t) M, (t)(f
) = |t=0 f ((t)).
t
n
n
Example 3.2. We have Tx R = R , where the isomorphism takes a Rn to the corresponding
velocity vector of the curve x + ta. That is,
n
X
f
ai
.
v(f ) = |t=0 f (x + ta) =
t
xi
i=1
S
Hence Tm M is a vector space of dimension n = dim M . The union T M = mM Tm M is a
vector bundle over M , called the tangent bundle. Coordinate charts for M give vector bundle
charts for T M . For a smooth map of manifolds : M M , the entirety of all maps dm
defines a smooth vector bundle map
d : T M T M .
A vector field on M is a derivation X : C (M ) C (M ). That is, it is a linear map
satisfying
X(f g) = X(f )g + f X(g).
The space of vector fields is denoted X(M ) = Der(C (M )). Vector fields are local, in the sense
that for any open subset U there is a well-defined restriction X|U X(U ) such that X|U (f |U ) =
(X(f ))|U . For any vector field, one obtains tangent vectors Xm Tm M by Xm (f ) = X(f )|m .
One can think of a vector field as an assignment of tangent vectors, depending smoothly on
m. More precisely, a vector field is a smooth
section of the tangent bundle T M . In local
P
for all f C (M ).
In short, X = Y where : C (N ) C (M ), f 7 f .
One has X Y if and only if Y(m) = dm (Xm ). From the definitions, one checks
X1 Y1 , X2 Y2 [X1 , X2 ] [Y1 , Y2 ].
Example 3.4. Let j : S M be an embedded submanifold. We say that a vector field X is
tangent to S if Xm Tm S Tm M for all m S. We claim that if two vector fields are tangent
to S then so is their Lie bracket. That is, the vector fields on M that are tangent to S form a
Lie subalgebra.
Indeed, the definition means that there exists a vector field XS X(S) such that XS j X.
Hence, if X, Y are tangent to S, then [XS , YS ] j [X, Y ], so [XS , YS ] is tangent.
Similarly, the vector fields vanishing on S are a Lie subalgebra.
Let X X(M ). A curve (t), t J R is called an integral curve of X if for all t J,
(t)
= X(t) .
In local coordinates, this is an ODE dxi = ai (x(t)). The existence and uniqueness theorem for
dt
ODEs (applied in coordinate charts, and then patching the local solutions) shows that for any
m M , there is a unique maximal integral curve (t), t Jm with (0) = m.
Definition 3.5. A vector field X is complete if for all m M , the maximal integral curve with
(0) = m is defined for all t R.
In this case, one obtains a smooth map
: R M M, (t, m) 7 t (m)
such that (t) = t (m) is the integral curve through m. The uniqueness property gives
0 = Id, t1 +t2 = t1 t2
i.e. t 7 t is a group homomorphism. Conversely, given such a group homomorphism such
that the map is smooth, one obtains a vector field X by setting
X = |t=0 t ,
t
|t=0 t .f =
|t=0 (1
X.f =
t ) f.
t
t
If t is a flow, we have 1
= t .
t
The Lie bracket of vector fields measure the non-commutativity of their flows. In particular,
Y
if X, Y are complete vector fields, with flows X
t , s , then [X, Y ] = 0 if and only if
Y
Y
X
X
t s = s t .
Y
In this case, X + Y is again a complete vector field with flow tX+Y = X
t t . (The right
hand side defines a flow since the flows of X, Y commute, and the corresponding vector field is
identified by taking a derivative at t = 0.)
3.2. The Lie algebra of a Lie group. Let G be a Lie group, and T G its tangent bundle.
For all a G, the left,right translations
La : G G, g 7 ag
Ra : G G, g 7 ga
are smooth maps. Their differentials at e define isomorphisms dg La : Tg G Tag G, and similarly for Ra . Let
g = Te G
be the tangent space to the group unit.
A vector field X X(G) is called left-invariant if
X La X
La .
10
L = |t=0 (t ) .
t
Letting (t) denote the unique integral curve with (0) = e. It has the property
(t1 + t2 ) = (t1 ) (t2 ),
and the flow of L is given by right translations:
t (g) = g (t).
Similarly, the right-invariant vector fields R are complete. (t) is an integral curve for R as
well, and the flow of R is given by left translations, g 7 (t)g.
11
Proof. If (t), t J R is an integral curve of a left-invariant vector field L , then its left
translates a(t) are again integral curves. In particular, for t0 J the curve t 7 (t0 )(t) is
again an integral curve. Hence it coincides with (t0 + t) for all t J (J t0 ). In this way,
an integral curve defined for small |t| can be extended to an integral curve for all t, i.e. L is
complete.
Since L is left-invariant, so is its flow t . Hence
t (g) = t Lg (e) = Lg t (e) = gt (e) = g (t).
The property t1 +t2 = t1 t2 shows that (t1 +t2 ) = (t1 ) (t2 ). Finally, since L Inv R ,
the image
Inv( (t)) = (t)1 = (t)
is an integral curve of R . Equivalently, (t) is an integral curve of R .
12
Proof. By the previous Proposition, (t) = t (1) = exp(t). For the second claim, note that
[, ] = 0 implies that L , L commute. Hence their flows t , t , and t t is the flow of
L + L . Hence it coincides with +
. Applying to e, we get (t) (t) = + (t). Now put
t
t = 1.
In terms of the exponential map, we may now write the flow of L as t (g) = g exp(t),
and similarly for the flow of R . That is,
L = |t=0 Rexp(t)
, R = |t=0 Lexp(t) .
t
t
Proposition 4.6. The exponential map is functorial with respect to Lie group homomorphisms
: G H. That is, we have a commutative diagram
G
x
exp
H
x
exp
g h
de
Proposition 4.7. Let G GL(n, R) be a matrix Lie group, and g gl(n, R) its Lie algebra.
Then exp : g G is just the exponential map for matrices,
X
1 n
.
exp() =
n!
n=0
Proof. By the previous Proposition, applied to the inclusion of G in GL(n, R), the exponential
map for G is just the restriction
for GL(n, R). Hence it suffices to prove the claim for
Pof that
tn n
G = GL(n, R). The function n=0 n! is a 1-parameter group in GL(n, R), with derivative
at 0 equal to gl(n, R). Hence it coincides with exp(t). Now put t = 1.
Proposition 4.8. For a matrix Lie group G GL(n, R), the Lie bracket on g = TI G is just
the commutator of matrices.
we have
Proof. It suffices to prove for G = GL(n, R). Using L = t
Rexp(t)
t=0
R
R
R
)
(R
s s=0 t t=0 exp(t) exp(s) exp(t) exp(s)
L Rexp(s)
L)
=
(R
s s=0 exp(s)
= L L L L
= [, ]L .
13
Rexp(t)
Rexp(s)
Rexp(t)
Rexp(s)
= Rexp(t)
exp(s) exp(t) exp(s) .
Since the Lie group exponential map for GL(n, R) coincides with the exponential map for
matrices, we may use Taylors expansion,
exp(t) exp(s) exp(t) exp(s) = I + st( ) + . . . = exp(st( )) + . . .
where . . . denotes terms that are cubic or higher in s, t. Hence
Rexp(t)
exp(s) exp(t) exp(s) = Rexp(st() + . . .
and consequently
=
= ( )L .
Rexp(t)
R
exp(s) exp(t) exp(s)
s s=0 t t=0
s s=0 t t=0 exp(st())
We conclude that [, ] = .
Remark 4.9. Had we defined the Lie algebra using right-invariant vector fields, we would have
obtained minus the commutator of matrices. Nonetheless, some authors use that convention.
The exponential map gives local coordinates for the group G on a neighborhood of e:
Proposition 4.10. The differential of the exponential map at the origin is d0 exp = id. As a
consequence, there is an open neighborhood U of 0 g such that the exponential map restricts
to a diffeomorphism U exp(U ).
Proof. Let (t) = t. Then (0)
14
Proof. The implication is obvious. For the opposite direction, suppose limn [vn ] = [v].
||v||
n ||
Let an N be defined by an 1 < ||v
an . Since vn 0, we have limn an ||v
||v|| = 1, and
n ||
||vn ||
an vn = an
[vn ] ||v|| [v] ||v|| = v.
||v||
Proof of E. Cartans theorem. It suffices to construct a submanifold chart near e H. (By
left translation, one then obtains submanifold charts near arbitrary a H.) Choose an inner
product on g.
We begin with a candidate for the Lie algebra of H. Let W g be the subset such that
W if and only if either = 0, or 6= 0 and there exists n 6= 0 with
exp(n ) H, n 0, [n ] [].
We will now show the following:
(i) exp(W ) H,
(ii) W is a subspace of g,
(iii) There is an open neighborhood U of 0 and a diffeomorphism : U (U ) G with
(0) = e such that
(U W ) = (U ) H.
(Thus defines a submanifold chart near e.)
Step (i). Let W \{0}, with sequence n as in the definition of W . By the Lemma, there are
an N with an n . Since exp(an n ) = exp(n )an H, and H is closed, it follows that
exp() = lim exp(an n ) H.
n
Step (ii). Since the subset W is invariant under scalar multiplication, we just have to show
that it is closed under addition. Suppose , W . To show that + W , we may assume
that , , + are all non-zero. For t sufficiently small, we have
exp(t) exp(t) = exp(u(t))
for some smooth curve t 7 u(t) g with u(0) = 0. Then exp(u(t)) H and
u(h)
1
= u(0)
= + .
lim n u( ) = lim
n
h0 h
n
hence u( n1 ) 0, exp(u( n1 ) H, [u( n1 )] [ + ]. This shows [ + ] W , proving (ii).
15
16
17
6.3. The adjoint representation of g. Let Der(g) be the Lie algebra of derivations of the
Lie algebra g. There is a Lie algebra morphism,
ad : g Der(g), 7 [, ].
The fact that ad is a derivation follows from the Jacobi identity; the fact that 7 ad it
is a Lie algebra morphism is again the Jacobi identity. The kernel of ad is the center of the
Lie algebra g, i.e. elements having zero bracket with all elements of g, while the image is
the Lie subalgebra Inn(g) Der(g) of inner derivations. It is a normal Lie subalgebra, i.e
[Der(g), Inn(g)] Inn(g), and the quotient Lie algebra Out(g) are the outer automorphims.
Suppose now that G is a Lie group, with Lie algebra g. We have remarked above that
the Lie algebra of Aut(g) is Der(g). Recall that the differential of any G-representation is a
g-representation. In particular, we can consider the differential of G Aut(g).
Theorem 6.4. If g is the Lie algebra of G, then the adjoint representation ad : g Der(g) is
the differential of the adjoint representation Ad : G Aut(g). One has the equality of operators
exp(ad ) = Ad(exp )
for all g.
Adexp(t) = ad . This is easy if G is a matrix
Proof. For the first part we have to show t
t=0
Lie group:
Ad
=
exp(t) exp(t) = = [, ].
exp(t)
t t=0
t t=0
For general Lie groups we compute, using
exp(s Adexp(t) ) = Adexp(t) exp(s) = exp(t) exp(s) exp(t),
(Adexp(t) )L =
R
t t=0
t t=0 s s=0 exp(s Adexp(t) )
=
R
t t=0 s s=0 exp(t) exp(s) exp(t)
R
R
=
R
t t=0 s s=0 exp(t) exp(s) exp(t)
= Rexp(t)
L Rexp(t)
t t=0
= [ L , L ]
= [, ]L = (ad )L .
This proves the first part. The second part is the commutativity of the diagram
Ad
G Aut(g)
x
x
exp
exp
g Der(g)
ad
which is just a special case of the functoriality property of exp with respect to Lie group
morphisms.
18
Remark 6.5. As a special case, this formula holds for matrices. That is, for B, C Matn (R),
eB C eB =
X
1
[B, [B, [B, C] ]].
n!
n=0
The formula also holds in some other contexts, e.g. if B, C are elements of an algebra with B
nilpotent (i.e. B N = 0 for some N ). In this case, both the exponential series for eB and the
series on the right hand side are finite. (Indeed, [B, [B, [B, C] ]] with n Bs is a sum of
terms B j CB nj , and hence must vanish if n 2N .)
Here the operator on the right hand side is defined to be the result of substituting ad in
z
the entire holomorphic function 1ez . Equivalently, it may be written as an integral
1 exp( ad )
=
ad
ds exp(s ad ).
Lexp()
ds (exp(s ad ))
(Lexp() (f ))(exp( + t)) = f (exp() exp( + t)).
t t=0
t t=0
19
R
=
R
t t=0 exp() exp(+t)
Rexp(s) Rexp(s(+t)
t
s
t=0
0
Z 1
(t)L Rexp(s(+t)
ds
=
R
t t=0 exp(s)
0
Z 1
ds Rexp(s)
L Rexp(s())
=
0
Z 1
ds (Adexp(s) )L
=
0
Z 1
ds (exp(s ad ))L .
=
1
ds
R1
0
Corollary 7.2. The exponential map is a local diffeomorphism near g if and only if ad
has no eigenvalue in the set 2iZ\{0}.
Proof. d exp is an isomorphism if and only if
1exp( ad )
ad
exp(f ) = de L1
exp(t) t exp(t) = .
t
On the other hand, by the formula for the differential of exp,
(de Lexp(f ) )1
(de Lexp(f ) )1
f
1 e adf f
exp(f ) = (de Lexp(f ) )1 (df exp)( ) =
( ).
t
t
adf
t
Hence
adf
df
.
=
dt
1 e adf
|
R
u u=0 exp((s+u))
R
= Rexp(s)
L =
s exp(s)
|
R
R
= Rexp(s) .
u u=0 exp(u) exp(s)
L Rexp(s)
for all g. Proof:
R
s exp(s)
20
X
log(w)
(1)n+1
=
1
+
(w 1)n ,
1 w1
n(n
+
1)
n=1
we may write the right hand side as (eadf ). By Applying Ad to the defining equation for f
we obtain eadf = ead et ad . Hence
df
= (ead et ad ).
dt
Finally, integrating from 0 to 1 and using f (0) = , f (1) = log(exp() exp()), we find:
Z 1
log(exp() exp()) = +
(ead et ad )dt .
0
w 1 = ead et ad 1 =
X tj
adi adj
i!j!
i+j1
in the power series expansion of , and integrates the resulting series in t. We arrive at:
Theorem 7.3 (Baker-Campbell-Hausdorff series). Let G be a Lie group, with exponential map
exp : g G. For , g sufficiently small we have the following formula
Z
n
X
(1)n+1 1 X tj
.
dt
adi adj
log(exp() exp()) = + +
n(n + 1) 0
i!j!
n=1
i+j1
An important point is that the resulting Taylor series in , is a Lie series: all terms of the
nr
2
series are of the form of a constant times adn 1 adm
ad . The first few terms read,
log(exp() exp()) = + + 12 [, ] +
1
1
1
[, [, ]] [, [, ]] + [, [, [, ]]] + . . . .
12
12
24
21
We will often write g.m rather than Ag (m). With this notation, g1 .(g2 .m) = (g1 g2 ).m and
e.m = m. A map : M1 M2 between G-manifolds is called G-equivariant if g.(m) =
(g.m) for all m M , i.e. the following diagram commutes:
G M1 M1
yid
y
G M2 M2
where the horizontal maps are the action maps.
Examples 8.2.
(a) An R-action on M is the same thing as a global flow.
(b) The group G acts M = G by right multiplication, Ag = Rg1 , left multiplication,
Ag = Lg , and by conjugation, Ag = Adg = Lg Rg1 . The left and right action
commute, hence they define an action of GG. The conjugation action can be regarded
as the action of the diagonal subgroup G G G.
(c) Any G-representation G End(V ) can be regarded as a G-action, by viewing V as a
manifold.
(d) For any closed subgroup H G, the space of right cosets G/H = {gH| g G}
has a unique manifold structure such that the quotient map G G/H is a smooth
submersion, and the action of G by left multiplication on G descends to a smooth
G-action on G/H. (Some ideas of the proof will be explained below.)
(e) The defining representation of the orthogonal group O(n) on Rn restricts to an action
on the unit sphere S n1 , which in turn descends to an action on the projective space
RP (n 1). One also has actions on the Grassmann manifold GrR (k, n) of k-planes in
Rn , on the flag manifold Fl(n) GrR (1, n) GrR (n 1, n) (consisting of sequences
of subspaces V1 Vn1 Rn with dim Vi = i), and various types of partial flag
manifolds. These examples are all of the form O(n)/H for various choices of H. (E.g,
for Gr(k, n) one takes H to be the subgroup preserving Rk Rn .)
8.2. Lie algebra actions.
Definition 8.3. An action of a finite-dimensional Lie algebra g on M is a Lie algebra homomorphism g X(M ), 7 A such that the action map
g M T M, (, m) 7 A |m
is smooth.
We will often write M =: A for the vector field corresponding to . Thus, [M , M ] = [, ]M
for all , g. A smooth map : M1 M2 between g-manifolds is called equivariant if
M1 M2 for all g, i.e. if the following diagram commutes
g M1 T M1
yid
yd
g M2 T M2
where the horizontal maps are the action maps.
Examples 8.4.
(a) Any vector field X defines an action of the Abelian Lie algebra R, by
7 X.
22
(b) Any Lie algebra representation : g gl(V ) may be viewed as a Lie algebra action
d
|t=0 f (v t()v) = hdv f, ()vi, f C (V )
dt
defines a g-action. Here dv f : Tv V R is viewed as an element of V . Using a basis ea
n
of V to identify
P V = bR , and introducing the components of g in the representation
as ().ea = b ()a eb the generating vector fields are
X
V =
()ba xa b .
x
(A f )(v) =
ab
Note that the components of the generating vector fields are homogeneous linear functions in x. Any g-action on V with this property comes from a linear g-representation.
(c) For any Lie group G, we have actions of its Lie algebra g by A = L , A = R and
A = L R .
(d) Given a closed subgroup H G, the vector fields R X(G), g are invariant
under the right multiplication, hence they are related under the quotient map to vector
fields on G/H. That is, there is a unique g-action on G/H such that the quotient map
G G/H is equivariant.
Definition 8.5. Let G be a Lie group with Lie algebra g. Given a G-action g 7 Ag on M , one
defines its generating vector fields by
d
A = Aexp(t) .
dt t=0
Example 8.6. The generating vector field for the action by right multiplication Aa = Ra1 are
the left-invariant vector fields,
= L.
A = |t=0 Rexp(t)
t
Similarly, the generating vector fields for the action by left multiplication Aa = La are R ,
and those for the conjugation action Ada = La Ra1 are L R .
Observe that if : M1 M2 is an equivariant map of G-manifolds, then the generating
vector fields for the action are -related.
Theorem 8.7. The generating vector fields of any G-action g Ag define a g-action A .
Proof. Write M := A for the generating vector fields of a G-action on M . We have to show
that 7 M is a Lie algebra morphism. Note that the action map
: G M M, (a, m) 7 a.m
is G-equivariant, relative to the given G-action on M and the action g.(a, m) = (ga, m) on
G M . Hence GM M . But GM = R (viewed as vector fields on the product
G M ), hence 7 GM is a Lie algebra morphism. It follows that
0 = [(1 )GM , (1 )GM ] [1 , 2 ]GM [(1 )M , (2 )M ] [1 , 2 ]M .
Since is a surjective submersion (i.e. the differential d : T (G M ) T M is surjective),
this shows that [(1 )M , (2 )M ] [1 , 2 ]M = 0.
23
8.3. Integrating Lie algebra actions. Let us now consider the inverse problem: For a Lie
group G with Lie algebra g, integrating a given g-action to a G-action. The construction will
use some facts about foliations.
Let M be a manifold. A rank k distribution on M is a C (M )-linear subspace R X(M )
of the space of vector fields, such that at any point m M , the subspace
Em = {Xm | X R}
is of dimension k. The subspaces Em define a rank k vector bundle E T M with R = (E),
hence a distribution is equivalently given by this subbundle E. An integral submanifold of the
distribution R is a k-dimensional submanifold S such that all X R are tangent to S. In terms
of E, this means that Tm S = Em for all m S. The distribution is called integrable if for all
m M there exists an integral submanifold containing m. In this case, there exists a maximal
such submanifold, Lm . The decomposition of M into maximal integral submanifolds is called
a k-dimensional foliation of M , the maximal integral submanifolds themselves are called the
leaves of the foliation.
Not every distribution is integrable. Recall that if two vector fields are tangent to a submanifold, then so is their Lie bracket. Hence, a necessary condition for integrability of a distribution
is that R is a Lie subalgebra. Frobenius theorem gives the converse:
Theorem 8.8 (Frobenius theorem). A rank k distribution R X(M ) is integrable if and only
if R is a Lie subalgebra.
The idea of proof is to show that if R is a Lie subalgebra, then the C (M )-module R is
spanned, near any m M , by k commuting vector fields. One then uses the flow of these
vector fields to construct integral submanifold.
Exercise 8.9. Prove Frobenius theorem for distributions R of rank k = 2. (Hint: If X R
with Xm 6= 0, one can choose local coordinates such that X = x 1 . Given a second vector field
Y R, such that [X, Y ] R and Xm , Ym are linearly independent, show that one can replace
Y by some Z = aX + bY R such that bm 6= 0 and [X, Z] = 0 on a neighborhood of m.)
Exercise 8.10. Give an example of a non-integrable rank 2 distribution on R3 .
Given a Lie algebra of dimension k and a free g-action on M (i.e. M |m = 0 implies = 0),
one obtains an integrable rank k distribution R as the span (over C (M )) of the M s. We
use this to prove:
Theorem 8.11. Let G be a connected, simply connected Lie group with Lie algebra g. A Lie
algebra action g X(M ),
7 M integrates to an action of G if and only if the vector fields
M are all complete.
Proof of the theorem. The idea of proof is to express the G-action in terms of a foliation.
Given a G-action on M , consider the diagonal G-action on G M , where G acts on itself by
left multiplication. The orbits of this action define a foliation of G M , with leaves indexed
by the elements of m:
Lm = {(g, g.m)| g G}.
Let pr1 , pr2 the projections from GM to the two factors. Then pr1 restricts to diffeomorphisms
m : Lm G, and we recover the action as
1
g.m = pr2 (m
(g)).
24
Given a g-action, our plan is to construct the foliation from an integrable distribution.
Let
7 M be a given g-action. Consider the diagonal g action on G M ,
GM = ( R , M ) X(G M ).
Note that the vector fields M
c are complete, since it is the sum of commuting vector fields,
R
both of which are complete. If t is the flow of M , the flow of M
c = ( , M ) is given by
b = (Lexp(t) , ) Diff(G M ).
t
t
mi = (1i 11 )(m)
t
t
t [0, 1]
connects (gi1 , mi1 ) to (gi , mi ), and stays within a leaf of the foliation (since it is given by the
flow). Hence, by concatenation we obtain a (piecewise smooth) path in Lm connecting (e, m)
1 (g) 6= .
to (gr , mr ) = (g, mr ). In particular, m
For any (g, x) Lm the tangent map d(g,x) m is an isomorphism. Hence m : Lm G is
a (surjective) covering map. Since G is simply connected by assumption, we conclude that
1 (g)). Concretely, the
m : Lm G is a diffeomorphism. We now define Ag (m) = pr2 (m
construction above shows that if g = exp(r ) exp(1 ) then
Ag (m) = (1r 11 )(m).
From this description it is clear that Agh = Ag Ah .
Let us remark that, in general, one cannot drop the assumption that G is simply connected.
Consider for example G = SU(2), with su(2)-action 7 R . This exponentiates to an action
of SU(2) by left multiplication. But su(2)
= so(3) as Lie algebras, and the so(3)-action does
not exponentiate to an action of the group SO(3).
As an important special case, we obtain:
Theorem 8.12. Let H, G be Lie groups, with Lie algebras h, g. If H is connected and simply connected, then any Lie algebra morphism : h g integrates uniquely to a Lie group
morphism : H G.
Proof. Define an h-action on G by 7 ()R . Since the right-invariant vector fields are
complete, this action integrates to a Lie group action A : H Diff(G). This action commutes
with the action of G by right multiplication. Hence, Ah (g) = (h)g where (h) = Ah (e). The
action property now shows (h1 )(h2 ) = (h1 h2 ), so that : H G is a Lie group morphism
integrating .
Corollary 8.13. Let G be a connected, simply connected Lie group, with Lie algebra g. Then
any g-representation on a finite-dimensional vector space V integrates to a G-representation
on V .
25
By Ados theorem, any finite-dimensional Lie algebra g is isomorphic to a matrix Lie algebra.
We will skip the proof of this important (but relatively deep) result, since it involves a considerable amount of structure theory of Lie algebras. Given such a presentation g gl(n, R),
the Lemma gives a Lie subgroup G GL(n, R) integrating g. Replacing G with its universal
covering, this proves:
Theorem 8.16 (Lies third theorem). For any finite-dimensional real Lie algebra g, there
exists a connected, simply connected Lie group G, unique up to isomorphism, having g as its
Lie algebra.
The book by Duistermaat-Kolk contains a different, more conceptual proof of Cartans theorem. This new proof has found important generalizations to the integration of Lie algebroids.
In conjunction with the previous Theorem, Lies third theorem gives an equivalence between
the categories of finite-dimensional Lie algebras g and connected, simply-connected Lie groups
G.
9. Universal covering groups
e
Given a connected topological space X with base point x0 , one defines the covering space X
as equivalence classes of paths : [0, 1] X with (0) = x0 . Here the equivalence is that of
e X. The
homotopy relative to fixed endpoints. The map taking [] to (1) is a covering p : X
covering space carries an action of the fundamental group 1 (X), given as equivalence classes
of paths with (1) = x0 , i.e. 1 (X) = p1 (x0 ). The group structure is given by concatenation
of paths
(
1 (2t)
0 t 21 ,
,
(1 2 )(t) =
2 (2t 1) 12 t 1
26
i.e. [1 ][2 ] = [1 2 ] (one shows that this is well-defined). If X = M is a manifold, then 1 (X)
e by deck transformations, this action is again induced by concatenation of paths:
acts on X
A[] ([]) = [ ].
e G
eG
e
]
^
Proof. The group multiplication and inversion lifts to smooth maps M
ult : G
G = G
f:G
e G.
e Using the functoriality properties of the universal covering construction, it
and Inv
e A proof that 1 (G) is central is outlined in
is clear that these define a group structure on G.
the following exercise.
Exercise 9.2. Recall that a subgroup H G is normal in G if Adg (H) H for all g G.
a) Let G be a connected Lie group, and H G a normal subgroup that is discrete (i.e.
0-dimensional). Show that H is a subgroup of the center of G.
b) Prove that the kernel of a Lie group morphism : G G is a closed normal subgroup.
The combination of these two facts shows that if a Lie group morphism is a covering, then
its kernel is a central subgroup.
Example 9.3. The universal covering group of the circle group G = U(1) is the additive group
R.
Example 9.4. SU(2) is the universal covering group of SO(3), and SU(2)SU(2) is the universal
covering group of SO(4). In both cases, the group of deck transformations is Z2 .
For all n 3, the fundamental group of SO(n) is Z2 . The universal cover is called the Spin
group and is denoted Spin(n). We have seen that Spin(3)
= SU(2) and Spin(4)
= SU(2)SU(2).
One can also show that Spin(5)
= Sp(2) and Spin(6) = SU(4). (See e.g. by lecture notes on
Lie groups and Clifford algebras, Section III.7.6.) Starting with n = 7, the spin groups are
new.
e of a Lie group G is compact if and
We will soon prove that the universal covering group G
only if G is compact with finite center.
e is central, and so G/
e is
If 1 (G) is any subgroup, then (viewed as a subgroup of G)
a Lie group covering G, with 1 (G)/ as its group of deck transformations.
27
28
one obtains a g-representation by restriction. That is, there is a 1-1 correspondence between
Lie algebra representations of g and algebra representations of U (g).
Let Cent(U (g)) be the center of the enveloping algebra. Given a g-representation : g
End(V ), the operators (x), x Cent(U (g)) commute with all (), g:
[(x), ()] = ([x, ]) = 0.
It follows that the eigenspaces of (x) for x Cent(U (g)) are g-invariant.
Exercise 10.4. Let g
= sl(2, R) be the Lie algebra with basis e, f, h and brackets [e, f ] =
h, [h, e] = 2e, [h, f ] = 2f . Show that
1
x = 2f e + h2 + h U (sl(2, R))
2
lies in the center of the enveloping algebra.
The construction of the enveloping algebra works for any Lie algebra, possibly of infinite
dimension. It is a non-trivial fact that the map j is always an inclusion. This is usually
obtained as a corollary to the Poincare-Birkhoff-Witt theorem. The statement of this Theorem
is as follows. Note that U (g) has a filtration
R = U (0) (g) U (1) (g) U (2) (g) ,
where U (k) (g) consists of linear combinations of products of at most k elements in g. That is,
L
U (k) (g) is the image of T (k) (g) = ik T i (g).
The filtration is compatible with the product, i.e. the product of an element of filtration
degree k with an element of filtration degree l has filtration degree k + l. Let
M
gr(U (g)) =
grk (U (g))
k=0
grk (U (g))
Lemma 10.5. The associated graded algebra gr(U (g)) is commutative. Hence, the map j : g
U (g) defines an algebra morphism
jS : S(g) gr(U (g))
Proof. If x = 1 k U (k) (g), and x = s(1) s(k) for some permutation s, then x x
U (k1) (g). (If s is a transposition of adjacent elements this is immediate from the definition;
but general permutations are products of such transpositions.) As a consequence, the products
of two elements of filtration degrees k, l is independent of their order modulo terms of filtration
degree k + l 1. Equivalently, the associated graded algebra is commutative.
Explicitly, the map is the direct sum over all
jS : S k (g) U (k) (g)/U (k1) (g), 1 k 7 1 k
Note that the map jS is surjective: Given y U (k) (g)/U (k1) (g), choose a lift y U (k) (g) given
as a linear combination of k-fold products of elements in g. The same linear combination, with
the product now interpreted in the symmetric algebra, defines an element x S k (g) with
jS (x) = y. The following important result states that jS is also injective.
29
for example,
sym(1 2 ) = 21 (1 2 + 2 1 ) = 1 2 12 [1 , 2 ].
Corollary 10.9. The symmetrization map sym : S(g) U (g) is an isomorphism of vector
spaces.
Another choice for f is to pick a basis e1 , . . . , en of g, and define f by
f (ei11 einn ) = ei11 einn .
Hence we obtain,
Corollary 10.10. If e1 , . . . , en is a basis of g, the products ei11 einn U (g) with ij 0 form
a basis of U (g).
Corollary 10.11. Suppose g1 , g2 are two Lie subalgebras of g such that g = g1 g2 as vector
spaces. Then the multiplication map
U (g1 ) U (g2 ) U (g)
is an isomorphism of vector spaces.
Indeed, the associated graded map is the multiplication S(g1 ) S(g2 ) S(g), which is
well-known to be an isomorphism. The following is left as an exercise:
Corollary 10.12. The algebra U (g) has no (left or right) zero divisors.
We will give a proof of the PBW theorem for the special case that g is the Lie algebra of
a Lie group G. (In particular, g is finite-dimensional.) The idea is to relate the enveloping
algebra to differential operators on G. For any manifold M , let
DO(k) (M ) = {D End(C (M ))| f0 , . . . , fk C (M ), adf0 adfk D = 0}
be the differential operators of degree k on M . Here adf = [f, ] is commutator with the
operator of multiplication by f . 4 By polarization, D DO(k) (M ) if and only if adk+1
D=0
f
for all f .
4In local coordinates, such operators are of the form
D=
X
i1 +...+in k
ai1 in
in
i1
xinn
xi11
30
31
Theorem 10.15. For any Lie group G, with Lie algebra g, the map 7 L extends to an
isomorphism
U (g) DOL (G)
where DOL (G) is the algebra of left-invariant differential operators on G.
11. Representation theory of sl(2, C).
11.1. Basic notions. Until now, we have mainly considered real Lie algebras. However, the
definition makes sense for any field, and in particular, we can consider complex Lie algebras.
Given a real Lie algebra g, its complexification g C is a complex Lie algebra. Consider in
particular the real Lie algebra su(n). Its complexification is the Lie algebra sl(n, C). Indeed,
sl(n, C) = su(n) isu(n)
is the decomposition of a trace-free complex matrix into its skew-adjoint and self-adjoint part.
Remark 11.1. Of course, sl(n, C) is also the complexification of sl(n, R). We have encountered
a similar phenomenon for the symplectic groups: The complexification of sp(n) is sp(n, C),
which is also the complexification of sp(n, R).
We will be interested in representations of Lie algebra g on complex vector spaces V , i.e.
Lie algebra morphisms g EndC (V ). Equivalently, this amounts to a morphism of complex
Lie algebras g C EndC (V ). If V is obtained by complexification of a real g-representation,
then V carries an g-equivariant conjugate linear complex conjugation map C : V V . Conversely, we may think of real g-representations as complex g-representations with the additional
structure of a g-equivariant conjugate linear automorphism of V .
Given a g-representation on V , one obtains representations on various associated spaces. For
instance, one has a representation on the symmetric power S k (V ) by
()(v1 vk ) =
k
X
v1 ()vj vk
k
X
v1 ()vj vk .
j=1
j=1
N
A similar formula gives a representation on the tensor powers k V , and both S k (V ), k (V )
Nk
(V ). One also obtains a dual representation on
are quotients of the representation on
V = Hom(V, C), by
h(), vi = h, ()(v)i,
V , v V.
32
(h)2 = 2
(f )1 = 2 ,
(f )2 = 0.
L
It extends to a representation on the symmetric algebra S(C2 ) = k0 S k (C2 ) by derivations,
preserving each of the summands S k (C2 ). Equivalently, this is the action on the space of
homogeneous polynomials of degree k on (C2 )
= C2 . Introduce the basis
vj =
1
j
kj
1 2 , j = 0, . . . , k
(k j)!j!
33
Since dim V < , there exists C such that V[] 6= 0 but V[+2] = 0. Pick a non-zero
v0 V[] , and put vj = j!1 (f )j v0 V[2j] , j = 0, 1, . . .. Then
(h)vj = ( 2j)vj , (f )vj = (j + 1)vj+1 .
We will show by induction that
(e)vj = ( + 1 j)vj1
with the convention v1 = 0. Indeed, if the formula holds for an index j 0 then
1
(e)(f )vj
(e)vj+1 =
j+1
1
=
(([e, f ])vj + (f )(e)vj )
j+1
1
=
((h)vj + ( + 1 j)(f )vj1 )
j+1
1
=
(( 2j)vj + ( + 1 j)j vj )
j+1
1
=
((j + 1) j j 2 )vj
j+1
= ( j)vj
which is the desired identity for j + 1. We see that the span of the vj is an invariant subspace, hence is all of V . Then non-zero vj are linearly independent (since they lie in different
eigenspaces for (h). Thus v0 , . . . , vk is a basis of V , where k = dim V 1. In particular,
vk+1 = 0. Putting j = k + 1 in the formula for (e)vj , we obtain 0 = ( k)vk , hence
= k.
Remark 11.4. For any complex number C, we obtain an infinite-dimensional representation
L() of sl(2, C) on span(w0 , w1 , w2 , . . .), by the formulas
(f )wj = (j + 1)wj+1 , (h)wj = ( 2j)wj ,
(e)wj = ( j + 1)wj1
34
Equivalently, the Theorem says that any invariant subspace V V has an invariant complement. We will first prove this result for the case that V has codimension 1 in V .
Lemma 11.8. Suppose V is a finite-dimensional sl(2, C)-representation, and V V an invariant subspace with dim(V /V ) = 1. Then V admits an invariant complement.
Proof. Suppose first that V admits a proper subrepresentation V1 V . Then V /V1 has
codimension 1 in V /V1 , hence by induction it admits an invariant complement W/V1 . Now
V1 has codimension 1 in W , hence, using induction again it admits an invariant complement
T W . Then T is an invariant complement to V in V .
We have thus reduced to the case that V has no proper subrepresentation, i.e. V is irreducible. There are now two subcases. Case 1: V
= V (0) is the trivial sl(2, C)-representation.
Since [sl(2, C), sl(2, C)] = sl(2, C) (each of the basis elements e, f, h may be written as a Lie
bracket), and sl(2, C).V V (because the action on the 1-dimensional space V /V is necessarily trivial), we have
sl(2, C).V = [sl(2, C), sl(2, C)].V sl(2, C).V = 0.
Hence sl(2, C) acts trivially, and the Lemma is obvious. Case 2: S = V (k) with k > 0. The
Casimir element Cas acts as k(k + 2)/2 > 0 on V (k). Hence the kernel of Cas on V is the
desired invariant complement.
Proof of Theorem 11.7. Suppose V V is an invariant subspace. Then Hom(V, V ) carries
an sl(2, C)-representation,
()(B) = [(), B]. Let W Hom(V, V ) be the subspace of
35
for any choice of n. (Given a Lie algebra morphism sl(2, R) gl(n, R), complexify to get a Lie
algebra morphism sl(2, C) gl(n, C).)
12. Compact Lie groups
In this section we will prove some basic facts about compact Lie groups G and their Lie
algebras g: (i) the existence of a bi-invariant positive measure, (ii) the existence of an invariant
inner product on g, (iii) the decomposition of g into center and simple ideals, (iv) the complete
reducibility of G-representations, (v) the surjectivity of the exponential map.
12.1. Modular function. For any Lie group G, one defines the modular function to be the
Lie group morphism
: G R , g 7 |detg(Adg )|.
Its differential is given by
de : g R, 7 trg(ad ),
|t=0 detg(Adexp(t) ) = t
|t=0 detg(exp(t ad )) = t
|t=0 exp(t trg(ad )) =
by the calculation t
trg(ad ). Here we have identified the Lie algebra of R with R, in such a way that the
exponential map is just the usual exponential of real numbers.
Lemma 12.1. For a compact Lie group, the modular function is trivial.
36
Proof. The range of the Lie group morphism G R , g 7 detg(Adg ) (as an image of a
compact set under a continuous map) is compact. But its easy to see that the only compact
subgroups of R are {1, 1} and {1}.
It also follows that, for G, compact, the infinitesimal modular function trg(ad ) is trivial.
Remark 12.2. A Lie group whose modular function is trivial is called unimodular. Besides
compact Lie groups, there are many other examples of unimodular lie groups. For instance, if
G is a connected Lie group whose Lie algebra
P is semi-simple, i.e. g = [g, g], then any g can
be written as a sum of commutators = i [i , i ]. But
X
X
ad =
ad[i ,i ] =
[adi , adi ]
i
has zero trace, since the trace vanishes on commutators. Similarly, if G is a connected Lie group
whose Lie algebra is nilpotent (i.e., the series g(0) = g, g(1) = [g, g], . . . , g(k+1) = [g, g(k) ], . . . is
eventually zero), then the operator ad is nilpotent (adN
= 0 for N sufficiently large). Hence
its eigenvalues are all 0, and consequently trg(ad ) = 0. An example is the Lie group of upper
triangular matrices with 1s on the diagonal.
An example of a Lie group that is not unimodular is the conformal group of the real line,
i.e. the 2-dimensional Lie group G of matrices of the form
t s
g=
0 1
with t > 0 and s R. In this example, one checks (g) = t.
12.2. Volume forms and densities. The modular function can also be interpreted in terms
of volume forms. We begin with a review of volume forms and densities on manifolds.
Definition 12.3. Let E be a vector space of dimension n. We define a vector space det(E )
consisting of maps : E E R satisfying
(Av1 , . . . , Avn ) = det(A) (v1 , . . . , vn ).
for all A GL(E) The non-zero elements of det(E ) are called volume forms on E. We also
define a space |det|(E ) of maps m : E E R satisfying
m(Av1 , . . . , Avn ) = | det(A)| m(v1 , . . . , vn )
for all A GL(E). The elements of |det|(E ) are called densities.
Both det(E ) and |det|(E ) are 1-dimensional vector spaces. Of course, det(E ) n E . A
volume form defines an orientation on E, where a basis v1 , . . . , vn is oriented if (v1 , . . . , vn ) >
0. It also defines a non-zero density m = || by putting ||(v1 , . . . , vn ) = |(v1 , . . . , vn )|.
Conversely, a positive density together with an orientation define a volume form. In fact, a
choice of orientation gives an isomorphism det(E )
= |det|(E ); a change of orientation changes
n
this isomorphism by a sign. The vector space R has a standard volume form 0 (taking the
oriented basis e1 , . . . , en ) to 1), hence a standard orientation and density |0 |. The latter is
typically denoted dn x, |dx| or similar. Given a linear map : E E , one obtains pull-back
maps : det((E ) ) det(E ) and : |det|((E ) ) |det|(E ); these are non-zero if and
only if is an isomorphism.
37
m.
comp (|det|(T M )) R, m 7
M
It is characterized as the unique linear map such that for any measure m supported in a chart
: U M , with U Rn one has
Z
Z
(x)|dx|
m=
M
5Note that measures in general are covariant objects: They push forward under continuous proper maps.
However, the push-forward of a smooth measure is not smooth, in general. Smooth measures (densities), on the
other hand, are contravariant objects.
38
Computing at the group unit e, we see that Ada1 e = det(Ada )1 e = e . The result for
densities is a consequence of that for volume forms.
Hence, if G is compact, any left-invariant density is also right-invariant. If G is compact
and connected, any left-invariantR volume form is also right-invariant. One can normalize the
left-invariant density such that G m = 1. The left-invariant measure on a Lie group G (not
necessarily normalized) is often denoted m = |dg|.
12.3. Basic properties of compact Lie groups. The existence of the bi-invariant measure
of finite integral lies at the heart of the theory of compact Lie groups. For instance, it implies
that the Lie algebra g of G admits an Ad-invariant inner product B: In fact, given an arbitrary
inner product B one may take B to be its G-average:
Z
1
B (Adg (), Adg ()) |dg|.
B(, ) =
vol(G) G
The Ad-invariance
(1)
follows from the bi-invariance if the measure. A symmetric bilinear form B on a Lie algebra g
is called ad-invariant if
(2)
B([, ], ) + B(, [, ]) = 0.
for all , , g. If g is the Lie algebra of a Lie group G, then any Ad-invariant bilinear form
is also ad-invariant, by differentiating the property (1).
As an application, we obtain the following decomposition of the Lie algebra of compact Lie
groups. An ideal in a Lie algebra g is a subspace h with [g, h] h. (In particular, h is a Lie
subalgebra). For instance, [g, g] an ideal. An ideal is the same thing as an invariant subspace
for the adjoint representation of g on itself. Note that for any two ideals h1 , h2 , their sum
h1 + h2 and their intersection h1 h2 are again ideals.
A Lie algebra is called simple if it is non-abelian and does not contain non-trivial ideals,
and semi-simple if it is a direct sum of simple ideals. For a simple Lie algebra, we must have
g = [g, g] (since [g, g] is a non-zero ideal), hence the same property g = [g, g] is also true for
semi-simple Lie algebras.
Proposition 12.5. The Lie algebra g of a compact Lie group G is a direct sum
g = z g1 gr ,
where z is the center of g, and the gi are simple ideals. One has [g, g] = g1 gr . The
decomposition is unique up to re-ordering of the summands.
Proof. Pick an invariant Euclidean inner product B on g. Then the orthogonal complement
(with respect to B) of any ideal h g is again an ideal. Indeed, [g, h] h implies
B([g, h ], h) = B(h , [g, h]) B(h , h) = 0,
hence [g, h ] h . Let z g be the center of g, and g = z . Then [g, g] = [g , g ] g . The
calculation
B([g, [g, g] ], g) = B([g, g] , [g, g]) = 0
39
shows [g, [g, g] ] = 0, hence [g, g] z = (g ) , i.e. g [g, g]. It follows that [g, g] = g .
This gives the desired decomposition g = z [g, g]. By a similar argument, we may inductively
decompose [g, g] into simple ideals. For the uniqueness part, suppose that h [g, g] is an
ideal not containing any of the gi . But then [gi , h] gi h = 0 for all i, which gives [g, h] =
L
i [gi , h] = 0. Hence h z.
Exercise 12.6. Show that for any Lie group G, the Lie algebra of the center of G is the center
of the Lie algebra.
12.5. The bi-invariant Riemannian metric. Recall some material from differential geometry. Suppose M is a manifold equipped with a pseudo-Riemannian metric B. That is, B
is a family of non-degenerate symmetric bilinear forms Bm : Tm M Tm M R depending
smoothly on m. A smooth curve : J M (with J R some interval) is called a geodesic if,
for any [t0 , t1 ] J, the restriction of is a critical point of the energy functional
Z t1
B((t),
(t))
dt.
E() =
t0
That is, for any variation of , given by a smooth 1-parameter family of curves s : [t0 , t1 ] M
(defined for small |s|), with 0 = and with fixed end points (s (t0 ) = (t0 ), s (t1 ) = (t1 ))
we have
E(s ) = 0.
s s=0
A geodesic is uniquely determined by its values (t ), (t
) at any point t J. It is one of
the consequences of the Hopf-Rinow theorem that if M is a compact, connected Riemannian
manifold, then any two points in M are joined by a length minimizing geodesic. The result
is false in general for pseudo-Riemannian metrics, and we will encounter a counterexample at
the end of this section.
Let G be a Lie group, with Lie algebra g. A non-degenerate symmetric bilinear form B : g
g R defines, via left translation, a left-invariant pseudo-Riemannian metric (still denoted B)
40
on G. If the bilinear form on g is Ad-invariant, then the pseudo-Riemannian metric on G is biinvariant. In particular, any compact Lie group admits a bi-invariant Riemannian metric. As
another example, the group GL(n, R) carries a bi-invariant pseudo-Riemannian metric defined
by the bilinear form B(1 , 2 ) = tr(1 2 ) on gl(n, R). It restricts to a pseudo-Riemannian metric
on SL(n, R).
Theorem 12.7. Let G be a Lie group with a bi-invariant pseudo-Riemannian metric B. Then
the geodesics on G are the left-translates (or right-translates) of the 1-parameter subgroups of
G.
Proof. Since B is bi-invariant, the left-translates or right-translates of geodesics are again
geodesics. Hence it suffices to consider geodesics (t) with (0) = e. For g, let (t) be the
unique geodesic with (0)
Notice
1 e ad(us )
|s=0
u s (t) =
|s=0 u s (t)
s
ad(us )
s
since u0 = 0, u 0 = 0. Hence, the s-derivative of E(s ) at s = 0 is
Z t1
B( u s (t), )
|s=0 E(s ) = 2
s
s s=0
t0
= 2B( us (t1 ), ) 2B( us (t0 ), ) = 0
s s=0
s s=0
Remark 12.8. A pseudo-Riemannian manifold is called geodesically complete if for any given
m M and v Tm M , the geodesic with (0) = m and (0)
41
An important result of Cartan says that a Lie algebra g is semi-simple (i.e. [g, g] = g) if
and only if its Killing form is non-degenerate (possibly indefinite). In this course, we will only
consider this statement for Lie algebra of compact Lie groups.
Proposition 12.12. Suppose g is the Lie algebra of a compact Lie group G. Then the Killing
form on g is negative semi-definite, with kernel the center z. Thus, if in addition g = [g, g] the
Killing form is negative definite.
Proof. Let B be an invariant inner product on g, i.e. B positive definite. The ad-invariance
says that ad is skew-symmetric relative to B. Hence it is diagonalizable (over C), and all
its eigenvalues are in iR. Consequently ad2 is symmetric relative to B, with non-positive
eigenvalues, and its kernel coincides with the kernel of ad . This shows that
(, ) = tr(ad2 ) 0,
with equality if and only if ad = 0, i.e. z.
12.7. Derivations. Let g be a Lie algebra. Recall that D End(g) is a derivation if and only
if D([, ]) = [D, ] + [, D] for all , g, that is
adD = [D, ad ].
Let Der(g) be the Lie algebra of derivations of a Lie algebra g, and Inn(g) the Lie subalgebra
of inner derivations, i.e. those of the form D = ad .
Proposition 12.13. Suppose the Killing form of g is non-degenerate. Then any derivation of
g is inner. In fact, Der(g) = Inn(g) = g.
6The Killing form is named after Wilhelm Killing. Killings contributions to Lie theory had long been under-
rated. In fact, he himself in 1880 had rediscovered Lie algebras independently of Lie (but about 10 years later).
In 1888 he had obtained the full classification of Lie algebra of compact Lie groups. Killings existence proofs
contained gaps, which were later filled by E. Cartan. The Cartan matrices, Cartan subalgebras, Weyl groups,
root systems Coxeter transformations etc. all appear in some form in W. Killings work (cf. Borel Essays in
the history of Lie groups and Lie algebras.) According A. J. Coleman (The greatest mathematical paper of all
time), he exhibited the characteristic equation of the Weyl group when Weyl was 3 years old and listed the
orders of the Coxeter transformation 19 years before Coxeter was born. On the other hand, the Killing form
was actually first considered by E. Cartan. Borel admits that he (Borel) was probably the first to use the term
Killing form.
42
Proof. Let D Der(g). Since the Killing form is non-degenerate, the exists g with
(, ) = tr(D ad )
for all g. The derivation D0 = D ad then satisfies tr(D0 ad ) = 0. For , g we
obtain
(D0 (), ) = tr(adD0 () ad ) = tr([D0 , ad ] ad ) = tr(D0 [ad , ad ]) = tr(D0 ad[,] ) = 0.
This shows D0 () = 0 for all , hence D0 = 0. By definition, Inn(g) is the image of the map
g Der(g), 7 ad . The kernel of this map is the center z of the Lie algebra. But if is
non-degenerate, the center z must be trivial.
If G is a Lie group with Lie algebra g, we had seen that Der(g) is the Lie algebra of the
Lie group Aut(g). The Proposition shows that if the Killing form is non-degenerate, then the
differential of the map G Aut(g) is an isomorphism. Hence, it defines a covering from the
identity component of G to the identity component of Aut(g).
Proposition 12.14. Suppose the Killing form on the finite-dimensional Lie algebra g is negative definite. Then g is the Lie algebra of a compact Lie group.
Proof. Since Aut(g) preserves the Killing form, we have
Aut(g) O(g, ),
the orthogonal group relative to . Since is negative definite, O(g, ) is compact. Hence
Aut(g) is a compact Lie group with Lie algebra Inn(g) = g.
Remark 12.15. The converse is true as well: That is, the Lie algebra of a Lie group G has
negative definite Killing form if and only if G is compact with finite center. The most common
proof of this fact is via a result that for a compact connected Lie group with finite center,
the fundamental group is finite. This result applies to the identity component of the group
Aut(g); hence the universal cover of the identity component of Aut(g) is compact. A different
proof, not using fundamental group calculations (but instead using some facts from Riemannian
geometry), may be found in Helgasons book Differential geometry, Lie groups and symmetric
spaces, Academic Press, page 133. We may get back to this later (if time allows).
13. The maximal torus of a compact Lie group
13.1. Abelian Lie groups. A Lie group G is called abelian if gh = hg for all g, h G, i.e. G
is equal to its center.7 A compact connected abelian group is called a torus. A Lie algebra g is
abelian (or commutative) if the Lie bracket is trivial, i.e. g equals its center.
Proposition 13.1. A connected Lie group G is abelian if and only if its Lie algebra g is
abelian. Furthermore, in this case the universal cover is
e=g
G
7Abelian groups are named after Nils Hendrik Abel. In the words of R. Bott, I could have come up with
that.
43
Proof. The Lie algebra g is abelian if and only if XL (G) is abelian, i.e. if and only if the flows
of any two left-invariant vector fields commute. Thus
exp() exp() = exp( + ) = exp() exp(),
for all , . Hence there is a neighborhood U of e such that any two elements in U commute.
Since any element of G is a product of elements in U , this is the case if and only if G is abelian.
We also see that in this case, exp : g G is a Lie group morphism. Its differential at 0 is the
identity, hence exp is a covering map. Since g is contractible, it is the universal cover of G.
We hence see that any abelian Lie group is of the form G = V /, where V
= g is a vector
space and is a discrete additive subgroup of V .
Lemma 13.2. There are linearly independent 1 , . . . , k V such that
= spanZ (1 , . . . , k ).
Proof. Choose any 1 such that Z1 = R1 . Suppose by induction that 1 , . . . , l are
linearly independent vectors such that spanZ (1 , . . . , l ) = spanR (1 , . . . , l ). If the Z-span is
all of , we are done. Otherwise, let V = V / spanR (1 , . . . , l ), = / spanZ (1 , . . . , l ), and
. Then span ( , . . . ,
= Rl+1
V satisfies Zl+1
pick l+1 such that its image l+1
l+1 ) =
Z 1
spanR (1 , . . . , l+1 ) .
Extending the i to a basis of V , thus identifying V = Rn , see that any abelian Lie group is
isomorphic to Rn /Zk for some n, k. That is:
Proposition 13.3. Any connected abelian Lie group is isomorphic to (R/Z)k Rl , for some
k, l. In particular, a k-dimensional torus is isomorphic to (R/Z)k .
For a torus T , we denote by t the kernel of the exponential map. We will call the
integral lattice. Thus
T = t/.
Definition 13.4. Let H be an abelian Lie group (possibly disconnected). An element h H is
called a topological generator of H if the subgroup {hk | k Z} generated by h is dense in T .
Theorem 13.5 (Kronecker lemma). Let u = (u1 , . . . , uk ) Rk , and t = exp(u) its image in
T = (R/Z)k . Then t is a topological generator if and only if 1, u1 , . . . , uk R are linearly
independent over the rationals Q. In particular, topological generators of tori exist.
Proof. Note that 1, u1 , . . . , uk R are linearly dependent over the rationals if and only if there
P
exist a1 , . . . , an , not all zero, such that ki=1 ai ui Z.
Let T = (R/Z)k , and let H be the closure of the subgroup generated by t. Then T /H is a
compact connected abelian group, i.e. is isomorphic to (R/Z)l for some l. If H 6= T , then l > 0,
and there exists a non-trivial group morphism T /H
= (R/Z)l R/Z (e.g. projection to the
first factor). By composition with the quotient map, it becomes a non-trivial group morphism
: T R/Z
44
that is trivial on H. Let d : Rk R be its differential, and put ai = d(ei ). Since d(Zk ) Z,
the ai are integers. Since vanishes on H, we have (t) = 0 mod Z, while
(t) = d(u)
=
k
X
ai ui
mod Z
mod Z.
i=1
P
P
That is, ki=1 ai ui Z. Conversely, if there are ai Z, not all zero, such that ki=1 ai ui Z,
Pk
define : T R/Z by (exp(v)) =
i=1 ai vi mod Z for v = (v1 , . . . , vl ). Then (t) = 0
mod Z so that is trivial on H. Since is not trivial on T , it follows that H is a proper
subgroup of T .
Let Aut(T ) be the group of automorphisms of the torus T . Any such automorphism induces
a Lie algebra automorphism d Aut(t) preserving . Conversely, given an automorphism of
the lattice , we obtain an automorphism of t = spanR () and hence of T = t/. That is,
Aut(T ) can be identified with the automorphisms of the lattice :
Aut(T ) = Aut().
After choosing an identification T = (R/Z)l , this is the discrete group GL(n, Z) = Matn (Z) of
matrices A with integer coefficients having an inverse with integer coefficients. By the formula
for the inverse matrix, this is the case if and only if the determinant is 1:
GL(n, Z) = {A Matn (Z)| det(A) 1}.
The group GL(n, Z) contains the semi-direct product (Z2 )n Sn , where Sn acts by permutation
of coordinates and (Z2 )n acts by sign changes. It is easy to see that the subgroup (Z2 )n Sn
is O(n, Z) = GL(n, Z) O(n), the transformations preserving also the metric. It is thus a
maximal compact subgroup of GL(n, Z).
13.2. Maximal tori. Let G be a compact Lie group, with Lie algebra g. A torus T G is
called a maximal torus if it is not properly contained in a larger subtorus of G.
Theorem 13.6. (E. Cartan) Let G be a compact, connected Lie group. Then any two maximal
tori of G are conjugate.
Proof. We have to show that is T, T G are two maximal tori, then there exists g G such
that Ada (T ) = T . Fix an invariant inner product B on g. Pick topological generators t, t of
T, T , and choose , g with exp() = t, exp( ) = t . Let a G be a group element for
which the function g 7 B( , Adg ()) takes on its maximum possible value. We will show that
Ada (T ) = T . To see this, let g. By definition of g, the function
t 7 B(Adexp(t) Ada (), )
takes on its maximum value at t = 0. Taking the derivative at t = 0, this gives
0 = B([, Ada ()], ) = B(, [Ada (), ]).
Since this is true for all , we obtain [ , Ada ()] = 0. Exponentiating , this shows Adt (Ada ()) =
Ada (). Exponentiating , it follows that Ada (t), t commute. Since these are generators, any
element in Ada (T ) commutes with any element in T . The group T Ada (T ) of products of
45
elements in T , Ada (T ) is connected and abelian, hence it is a torus. Since T , Ada (T ) are
maximal tori, we conclude T = T Ada (T ) = Ada (T ).
Definition 13.7. The rank l of a compact, connected Lie group is the dimension of its maximal
torus.
For example, U(n) has maximal torus given by diagonal matrices. Its rank is thus l = n.
We will discuss the maximal tori of the classical groups further below.
Exercise 13.8. The group SU(2) has maximal torus T the set of diagonal matrices diag(z, z 1 ).
Another natural choice of a maximal torus is T = SO(2) SU(2). Find all elements a G
such that Ada (T ) = T .
The Lie algebra t of a maximal torus T is a maximal abelian subalgebra of g, where a subalgebra is called abelian if it is commutative. Conversely, for any maximal abelian subalgebra the
subgroup exp(t) is automatically closed, hence is a maximal torus. Cartans theorem implies
that any two maximal abelian subalgebras t, t are conjugate under the adjoint representation.
That is, there exists a G such that Ada (t) = t .
Proposition 13.9 (Properties of maximal tori).
(a) Any element of a Lie group is contained in some maximal torus. That is, if T is a fixed maximal torus then
[
Ada (T ) = G.
aG
On the other hand, the intersection of all maximal tori is the center of G:
\
Ada (T ) = Z(G).
aG
Proof.
kZ
Let B0 be the identity component, which is thus a torus, and let m 0 be the smallest
number with gm B0 . Then B has connected components B0 , gB0 , . . . , gm1 B0 . The
element gm B0 can be written in the form km with k B0 . Thus h = gk1 gB0
satisfies hm = e. It follows that h generates a subgroup isomorphic to Zm , and the
product map B0 Zm B, (t, hi ) 7 thi is an isomorphism.
Pick a topological generator b B0 of the torus B0 . Then bm is again a topological
generator of B0 (by Kroneckers Lemma). Thus bh is a topological generator of B. But
bh is contained in some maximal torus T . Hence B T .
46
(c) By (b) there exists a maximal torus T containing T and g. But T already is a maximal
torus. Hence g T = T .
Exercise 13.10. Show that the subgroup of diagonal matrices in SO(n), n 3 is maximal
abelian. Since this is a discrete subgroup, this illustrates that maximal abelian subgroups need
not be maximal tori.
Proposition 13.11. dim(G/T ) is even.
Proof. Fix an invariant inner product on g. Since G is connected, the adjoint representation
takes values in SO(g). The action of T G fixes t, hence it restricts to a representation
T SO(t )
8where t
= g/t is the orthogonal complement with respect to B. Let t T be a topological
generator. Then Ad(t)|t has no eigenvalue 1. But any special orthogonal transformation on an
odd-dimensional Euclidean vector space fixes at least one vector. (Exercise.) Hence dim(g/t)
is even.
13.3. The Weyl group. For any subset S G of a Lie group, one defines its normalizer
N (S) (sometimes written NG (S) for clarity) to be the group of elements g G such that
Adg (S) S. If H is a closed subgroup of G, then N (H) is a closed subgroup. Since H is a
normal subgroup of N (H), the quotient N (H)/H inherits a group structure.
We are mainly interested in the normalizer of T . Thus, N (T ) is the stabilizer of T for the
G-action on the set of maximal tori. By Cartans theorem, this action is transitive, hence
the quotient space G/N (T ) is identified with the set of maximal tori. The adjoint action of
T N (T ) on T is of course trivial, but there is a non-trivial action of the quotient N (T )/T .
Definition 13.12. Let G be a compact, connected Lie group with maximal torus T . The quotient
W = NG (T )/T
is called the Weyl group of G relative to T .
Since any two maximal tori are conjugate, the Weyl groups are independent of T up to
isomorphism. More precisely, if T, T are two maximal tori, and a G with T = Ada (T ),
then N (T ) = Ada (N (T )), and hence Ada defines an isomorphism W W . There are many
natural actions of the Weyl group:
(a) W acts on T . This action is induced by the conjugation action of N (T ) on T (since
T N (T ) acts trivially). Note that this action on T is by Lie group automorphisms.
(b) W acts on t. This action is induced by the adjoint representation of N (T ) G on T
(since T N (T ) acts trivially). Of course, the action on t is just the differential of the
action on T .
(c) W acts on the lattice , the kernel of the exponential map exp : T . Indeed, exp : t
T is an N (T )-equivariant, hence W -equivariant, group morphism. Thus its kernel is a
W -invariant subset of t.
(d) W acts on G/T . The action of N (T ) on G by multiplication from the right (i.e n.g =
gn1 ) descends to a well-defined action of W = N (T )/T on G/T :
w.(gT ) = gn1 T
47
where n N (T ) represents w. Note that this action is free, that is, all stabilizer groups
are trivial. The quotient of the W -action on G/T is G/N (T ), the space of maximal tori
of G.
Example 13.13. For G = SU(2), with maximal torus T consisting of diagonal matrices, we have
N (T ) = T nT where
0 1
n=
.
1 0
Thus W = N (T )/T = Z2 , with n descending to the non-trivial generator. One easily checks
that the conjugation action of n on T permutes the two diagonal entries. The action on
t is given by reflection, 7 . The action on S 2 = G/T is the antipodal map, hence
(G/T )/W = G/N (T )
= RP (2).
Example 13.14. Let G = SO(3), with maximal torus given by rotations about the 3-axis. Thus,
T consists of matrices
cos() sin() 0
g() = sin() cos() 0 .
0
0
1
The normalizer N (T ) consist of all rotations preserving the 3-axis. The induced action on the
3-axis preserves the inner product, hence it is either trivial or the reflection. Elements in N (T )
fixing the axis are exactly the elements of T itself. The elements in N (T ) reversing the axis
are the rotations by about any axis orthogonal to the 3-axis. Thus W = Z2 .
Theorem 13.15. The Weyl group W of a compact, connected group G is a finite group.
Proof. We have to show that the identity component N (T )0 is T . Clearly, T N (T )0 . For
the other direction, consider the adjoint representation of N (T ) on t. As mentioned above this
action preserves the lattice . Since is discrete, the identity component N (T )0 acts trivially
on . It follows that N (T )0 acts trivially on t = spanR () and hence also on T = exp(t). That
is, N (T )0 ZG (T ) = T .
Proposition 13.16. The action of W on T (and likewise the action on t, ) is faithful. That
is, if a Weyl group element w acts trivially then w = 1.
Proof. If w acts trivially on , then it also acts trivially on t = spanR (). If w acts trivially on
t, then it also acts trivially on T = exp(t). If w acts trivially on T , then any element n N (T )
representing w lies in Z(T ) = T , thus w = 1.
48
13.4. Maximal tori and Weyl groups for the classical groups. We will now describe the
maximal tori and the Weyl groups for the classical groups. Recall that if T is a maximal torus,
then the Weyl group action
W Aut(T )
= Aut()
is by automorphism. Since W is finite, its image must lie in a compact subgroup of T . Recall
also that for the standard torus (R/Z)l , a maximal compact subgroup of Aut((R/Z)l ) is
O(l, Z) = (Z2 )l Sl GL(l, Z) = Aut((R/Z)l ).
To compute the Weyl group in the following examples of matrix Lie groups, we take into
account that the Weyl group action must preserve the set of eigenvalues of matrices t T .
13.4.1. The unitary and special unitary groups. For G = U(n), the diagonal matrices
z1
0
0 0
0
z2
0 0
diag(z1 , . . . , zn ) =
0
0
0 zn
R(1 )
0
0
0
R(
)
0
2
t(1 , . . . , m ) =
0
0
0
R(m )
49
permutation of the i , and (Z2 )m acts by sign changes. That is, we have an injective group
morphism
W (Z2 )m Sm .
To describe its image, let m (Z2 )m be the kernel of the product map (Z2 )m Z2 , corresponding to an even number of sign changes.
Theorem 13.18. The Weyl group W of SO(2m) is the semi-direct product m Sm .
Proof. The matrix g SO(2m), written in block form with 2 2-blocks, with entries gij =
gji = I, gkk = I for k 6= i, j, and all other blocks equal to zero, lies in N (T ). The corresponding Weyl group element permutes the i-th and j-th blocks of any t T . Hence Sn W . Next,
observe that
1 0
K=
O(2).
0 1
satisfies KR()R1 = R(). The block diagonal matrix, with blocks K in the i-th and jth diagonal entries, and identity matrices for the other diagonal entries, lies in N (T ) and its
action on T changes R(i ), R(j ) to R(i ), R(j ). Hence, we obtain all even numbers of sign
changes, confirming n W . It remains to show that the transformation t(1 , 2 , . . . , m ) 7
t(1 , 2 . . . , m ) does not lie in W . Suppose g N (T ) realizes this transformation, so that
gt(1 , 2 , . . . , m )g1 = t(1 , 2 , . . . , m ). As above, we find R(i )gij R(j ) = gij for i 2,
and R(1 )g1j = g1j R(j ) for i = 1, for all 1 , . . . , m . As before, we deduce from these
equations that g most be block diagonal. Thus g (O(2) O(2)) SO(2m). From
R(i )gii R(i ) = gii for i 2 we obtain gii SO(2) for i > 1. Since det(g) = 1, this forces
g11 SO(2), which however is incompatible with R(1 )g11 = g11 R(1 ).
13.4.3. The special orthogonal groups SO(2m + 1). Define an inclusion
j : O(2m) SO(2m + 1),
placing a given orthogonal matrix A in the upper left corner and det(A) in the lower left corner.
Let T be the standard maximal torus for SO(2m), and N (T ) its normalizer. Then T = j(T )
is a maximal torus for SO(2m + 1). The proof that T is maximal is essentially the same as for
SO(2m).
Theorem 13.19. The Weyl group of SO(2m + 1) is the semi-direct product (Z2 )m Sm .
Proof. As in the case of SO(2m), we see that the Weyl group must be a subgroup of (Z2 )m Sm .
Since j(N (T )) N (T ), we have an inclusion of Weyl groups W = m Sm W . Hence we
only need to show that the first Z2 is contained in W . The block diagonal matrix g O(2m)
with entries K, I, . . . , I down the diagonal satisfies gt(1 , . . . , m )g1 = t(1 , . . . , m ). Hence
j(g) N (T ) represents a generator of the Z2 .
13.4.4. The symplectic groups. Recall that Sp(n) is the subgroup of Matn (H) preserving the
norm on Hn . Alternatively, using the identification H = C2 , one can realize Sp(n) as a subgroup
of U(2n), consisting of matrices of the form
A B
(3)
B A
50
rank
name
l 1 SU(l + 1)
l 2 Spin(2l + 1)
l3
Sp(l)
l4
Spin(2l)
dim
W
+ 2l
Sl+1
2
2l + l (Z2 )l Sl
2l2 + l (Z2 )l Sl
2l2 l (Z2 )l1 Sl
l2
In the last row, (Z2 )l1 is viewed as the subgroup of (Z2 )l of tuples with product equal to 1.
Remarks 13.21.
(a) Note that the groups Sp(l) and Spin(2l + 1) have the same rank and
dimension, and isomorphic Weyl groups.
(b) For rank l = 1, Sp(1)
= SU(2)
= Spin(3). For rank l = 2, it is still true that Sp(2)
=
Spin(5). But for l > 2 the two groups Spin(2l+1), Sp(l) are non-isomorphic. To exclude
such coincidences, and to exclude the non-simple Lie groups Spin(4) = SU(2) SU(2),
one restricts the range of l as indicated above.
(c) As we will discuss later, the table is a complete list of simple, simply connected compact
Lie groups, with the exception of five aptly named exceptional Lie groups F4 , G2 , E6 , E7 , E8
that are more complicated to describe.
51
52
b = Hom(, U(1)).
with the finite group
14.2. Schurs Lemma. To proceed, we need the following simple but important fact.
Lemma 14.7 (Schur Lemma). Let G be any group, and : G GL(V ) a finite-dimensional
irreducible complex representation.
(a) If A End(V ) commutes with all (g), then A is a multiple of the identity matrix.
(b) If V is another finite-dimensional irreducible G-representation, then
(
1 if V
=V
dim(HomG (V, V )) =
0 otherwise
Proof. a) Let be an eigenvalue of A. Since ker(A ) is G-invariant, it must be all of V .
Hence A = I. b) For any G-equivariant map A : V V , the kernel and range of A are
sub-representations. Hence A = 0 or A is an isomorphism. If V, V are non-isomorphic, A
cannot be an isomorphism, so A = 0. If V, V are isomorphic, so that we might as well assume
V = V , b) follows from a).
For any two complex G-representations V, W , one calls HomG (V, W ) the space of intertwining
operators. If V is irreducible, and the representation W is completely reducible (as is automatic
for G a compact Lie group), the dimension dim HomG (V, W ) is the multiplicity of V in W .
The range of the map
HomG (V, W ) V W, A v 7 A(v)
is the V -isotypical subspace of W , i.e. the sum of all irreducible components isomorphic to V .
14.3. Weights of T -representations. For any X (T ), let C denote the T -representation
on C, with T acting via the homomorphism : T U(1).
Proposition 14.8. Any finite-dimensional irreducible representation of T is isomorphic to
C , for a unique weight X (T ). Thus, X (T ) labels the isomorphism classes of finitedimensional irreducible T -representations.
Proof. Let : T GL(V ) be irreducible. Since T is abelian, Schurs lemma shows that all
(t) act by scalars. Hence any v V spans an invariant subspace. Since V is irreducible,
it follows that dim V = 1, and the basis vector v gives an isomorphism V
= C. The image
(T ) GL(V ) = GL(1, C) is a compact subgroup, hence it must lie in U(1). Thus, becomes
a morphism : T U(1).
53
Any
Lfinite-dimensional complex T -representation V has a unique direct sum decomposition
V = X (T ) V , where the V are the C -isotypical subspaces. Thus V is the subspace on
which elements t T act as scalar multiplication by (t). Note that since dim C = 1, the
dimension of the space of intertwining operators coincides with the dimension of V . This is
called the multiplicity of the weight in V . We say that X (T ) is a weight of V if V 6= 0,
i.e. if the multiplicity is > 0, in this case V is called a weight space.
Let (V ) X (T ) be the set of all weights of the representation V . Then
M
V =
V .
(V )
54
55
Example 14.14. For G = SU(n), let T be the maximal torus given by diagonal matrices. Let T
be the maximal torus of U(n), again consisting of the diagonal matrices. Then X (T ) X (T ).
In terms of the standard basis 1 , . . . , n of X (T ), the lattice X (T ) has basis 1 2 , 2
3 , . . . , n1 n . The kernel of the projection map X (T ) X (T ) is a rank 1 lattice
generated by 1 + . . . + n . Thus, we can think of X (T ) as a quotient lattice
X (T ) = spanZ (1 , . . . , n )/ spanZ (1 + . . . + n ).
The images of i j under the quotient map are then the roots of SU(n). The root vectors
are the same as for U(n) (since they all lie in sl(n, C)).
One can get a picture of the rot system by identifying X (T ) with the orthogonal projection
of X (T ) to the subspace orthogonal to 1 + . . . + n , using the standard inner product on
X (T ) Z R = Rn . Note that the standard inner product is W = Sn -invariant, hence this
identification respects the Weyl group action. The projections of the i are
1
i = i (1 + . . . + n ), i = 1, . . . , n.
n
1
n1
Then , . . . ,
are a lattice basis of X (T ), and n = ( 1 + . . . + n1 ). The roots are
i
j
for i 6= j. Here is a picture of the weights and roots of SU(3):
PICTURE
block diagonal matrices t(1 , . . . , m ). Let X (T ) be the standard basis of the weight lattice. Thus j (t(1 , . . . , m )) = eij . The complexified Lie algebra gC = so(2m)C =: so(2m, C)
consists of skew-symmetric complex matrices. To find the root vectors, write the elements
T . Conjugation by t( , . . . , )
X so(2m, C) in block form, with 2 2-blocks Xij = Xji
1
m
changes the (i, j)-block to R(i )Xij R(j ). Let v+ , v C2 be non-zero column vectors with
R()v+ = ei v+ , R()v = ei v .
For instance, we may take
v+ =
T Mat (C) satisfy
The matrices v v
2
i
1
, v =
1
i
T
)R() = ei(+) ,
R()(v+ v+
T
)R() = ei(+) ,
R()(v v+
T
R()(v v+
)R() = ei(+) ,
T
R()(v v
)R() = ei() .
T in the (i, j) position, and its negative transpose in the (j, i)
Let i < j be given. Putting v v
position, and letting all other entries of X be zero, we obtain root vectors for the roots i j .
To summarize, SO(2m) has 2m(m 1) roots
R = {i j , i < j}.
This checks with dimensions, since dim T = m, dim SO(2m) = 2m2 m, so dim SO(2m)/T =
2(m2 m).
56
PICTURE
Example 14.16. Let G = SO(2m + 1). We write matrices in block form, corresponding to the
decomposition R2m+1 = R2 R2 R. Thus, X Mat2m+1 (C) has 2 2-blocks Xij for
i, j m, a 1 1-block Xm+1,m+1 , 2 1-blocks Xi,m+1 for i m, and 1 2-blocks Xm+1,i for
i m. As we saw earlier, the inclusion SO(2m) SO(2m+1) defines an isomorphism from the
maximal torus T of SO(2m) to a maximal torus T of SO(2m + 1). The latter consists of block
diagonal matrices, with 2 2-blocks gii = R(i ) for i = 1, . . . , m and 1 1-block gm+1,m+1 = 1.
Under the inclusion so(2m, C) so(2m+1, C), root vectors for the former become root vectors
for the latter. Hence, all i j are roots, as before. Additional root vectors X are obtained
by putting v as the Xi,m+1 block and its negative transpose in the Xm+1,i block, and letting
all other entries be zero. The corresponding roots are i . In summary, SO(2m + 1) has roots
R = {i j , 1 i < j m} {i , i = 1, . . . , m}.
PICTURE
This checks with dimensions: We have found 2m(m 1) + 2m = 2m2 roots, while dim SO(2m +
1)/T = (2m2 + m) m = 2m2 . Note that in this picture, the root system for SO(2m + 1)
naturally contains that for SO(2m). Note also the invariance under the Weyl group action in
both cases.
Example 14.17. Let G = Sp(n), viewed as the matrices in SU(2n) of the form
A B
,
B A
and let T be its standard maximal torus consisting of the diagonal matrices
Z 0
t=
0 Z
To find the roots, recall that the Lie algebra sp(n) consists of complex matrices of the form
a b
=
,
b a
with bT = b, cT = c. Thus
tt
ZaZ 1
ZbZ
Z 1 cZ 1 Z 1 aT Z
57
- Taking a = 0, c = 0 and letting b be a matrix having 1 in the (i, j) slot and zeroes
elsewhere, we obtain a root vector for the root i + j .
- Letting a = 0, b = 0, and letting c be a matrix having 1 in the (i, j) slot and zeroes
elsewhere, we obtain a root vector for the root i j .
- Letting b = 0, c = 0 and taking for a the matrix having aij = 1 has its only non-zero
entry, we obtain a root vector for i j (provided i 6= j).
Hence we have found
n(n + 1) n(n + 1)
+
+ (n2 n) = 2n2
2
2
roots:
R = {i j |1 i < j m} {2i | i = 1, . . . , m}
PICTURE
This checks with dimensions: dim(Sp(n)/T ) = (2n2 + n) n = 2n2 . Observe that the inclusion
u(n) sp(n) takes the root spaces of U(n) to root spaces of Sp(n). Hence, the set of roots of
U(n) is naturally a subset of the set of roots of Sp(n).
Suppose G, G are compact, connected Lie groups, and : G G is a covering map, with
kernel . Then restricts to a covering of the maximal tori, 1 T T 1. hence
X (T ) is a sublattice of X (T ), with quotient , while X (T ) is a sublattice of X (T ), with
b The roots of G are identified with the roots of G under the inclusion X (T )
quotient .
X (T ).
Example 14.18. Let G = SO(2m), and G = Spin(2m) its double cover. Let 1 , . . . , m be the
standard basis of the maximal torus T
= U(1)m . Each i : U(1) T may be regarded as
a loop in SO(2m), and in fact any of these represents a generator 1 (SO(2m)) = Z2 . With
a little work, P
one may thus show that X (T ) is the sublattice
Pm of X (T ) consisting of linear
m
i
combinations i=1 ai with integer coefficients, such that i=1 ai is even. Generators for this
lattice are, for example, 1 2 , 2 3 , . . . , n1 n , n1 + n . Dually, X (T ) is a lattice
containing X (T ) = spanZ (1 , . . . , m ) as a sublattice. It is generated by X (T ) together with
the vector 21 (1 + . . . + n ). The discussion for Spin(2m + 1) is similar. Here is, for example, a
picture of the root system for Spin(5):
PICTURE
15.1. First properties. We have already seen that the set of roots is W -invariant, and that
the roots come in pairs , with complex conjugate root spaces g = g . Another simple
property is
Proposition 15.1. For all , (gC ) = R {0},
[g , g ] = g+
58
Let us fix a non-degenerate Ad-invariant inner product B on g. Its restriction to t is a W invariant inner product on t. We use the same notation B for its extension to a non-degenerate
symmetric complex-bilinear form on gC , respectively tC .
Proposition 15.2. The spaces g , g for + 6= 0 are B-orthogonal, while g , g are nonsingularly paired.
Proof. If X g , X g , then
B(X , X ) = B(Ad(t)X , Ad(t)X ) = ( + )(t) B(X , X ),
hence + = 0 if B(X , X ) 6= 0.
15.2. The Lie subalgebras sl(2, C) gC , su(2) g. To proceed, we need to use some
representation theory of sl(2, C). Recall that sl(2, C) has the standard basis e, f, h with bracket
relations [e, f ] = h, [h, e] = 2e, [h, f ] = 2f . Any finite-dimensional sl(2, C)-representation is
a direct sum of irreducible representations. Furthermore, we had an explicit description of the
irreducible sl(2, C) representations. From this we read off:
Lemma 15.3. Let : sl(2, C) End(V ) be a finite-dimensional complex sl(2, C)-representation.
Then (h) has integer eigenvalues, and V is a direct sum of the eigenspaces Vm = ker((h)m).
For m > 0, the operator (f ) gives an injective map
(f ) : Vm Vm2 .
For m < 0, the operator (e) gives an injective map
(e) : Vm Vm+2 .
One has direct sum decompositions
V = ker(e) ran(f ) = ker(f ) ran(e).
Proof. All these claims are evident for irreducible representations V (k), hence they also hold
for direct sums of irreducibles.
The Lie algebra su(2) can be regarded as fixed point set of the conjugate linear involution
of sl(2, C), given by
h 7 h, e 7 f, f 7 e.
Indeed, the fixed point set of this involution is spanned by
X = i(e + f ), Y = f e, Z = i h,
59
and these vectors satisfy [X, Y ] = 2Z, [Y, Z] = 2X, [Z, X] = 2Y . In the matrix representation,
0 i
0 1
i 0
X=
, Y =
, Z=
.
i 0
1 0
0 i
[e , e ] = B(e , e )H ,
60
(5)
But T
= U(1), by exponentiating the isomorphism t u(1) = iR, ish 7 is. Hence
X (T ) = X (U(1)) = Z.
Definition 15.6. The co-root X (T ) corresponding to a root is the image of 1 Z
=
X (T ) under the inclusion (5). The set of co-roots is denoted R X (T ).
Note that 2ih t generates the integral lattice of T . Thus, corresponds to h
under the identification X (T ) Z R = it. That is,
d(h ) = h , i
for all X (T ).
Remark 15.7.
61
Example 15.8. Recall that SO(2m) has roots = i j for i 6= j, together with roots = i .
In terms of the standard inner product (, ) on X (T ) Z R, the co-roots for roots of the first
type are = i j , while for the second type we get = 2i . Note that these co-roots
for SO(2m) are precisely the roots for Sp(m). This is an example of Langlands duality.
15.4. Root lengths and angles. Choose a W -invariant inner product (, ) on E = X (T ) Z
R.
Proposition 15.9. Let , R be two roots, with |||| ||||. Suppose the angle between
, is not a multiple of 2 . Then one of the following three cases holds true:
||||2
= 1,
||||2
||||2
= 2,
||||2
||||2
= 3,
||||2
=
4
=
6
=
mod ,
mod ,
mod .
These properties of the root systems are nicely illustrated for the classical groups. Let us
also note the following consequence of this discussion:
Lemma 15.10. For all roots , R, the integer h , i lies in the interval [3, 3].
62
Lp
The direct sum
j=q g+j is an irreducible sl(2, C) -representation of dimension
p + q + 1.
(c) If , , + are all roots, then
[g , g ] = g+ .
Proof. We will regard g as an sl(2, C) -representation. By definition of the co-roots, we have
ad(h )e = h , ie
for e g .
(a) Suppose 6= is a root with h , i < 0. Since ad(h ) acts on g as a negative
scalar h , i < 0, the sl(2, C)-representation theory shows that ad(e ) : g g+ is
injective. In particular, g+ is non-zero.
(b) Suppose : sl(2, C) End(V ) is a finite-dimensional complex sl(2, C)-representation.
By the representation theory of sl(2, C), the number of odd-dimensional irreducible components occurring in V is equal to dim ker((h)), while the number of odd-dimensional
L
irreducible components is equal to dim ker((h) 1). Apply this to V = jZ g+j as
an sl(2, C) -representation. ad(h ) acts on g+j as h , i + 2j. The eigenvalues on V
are hence
L either all even or all odd, and their multiplicities are all one. This shows that
V = jZ g+j is irreducible. Let q, p be the largest integers such that g+p 6= 0,
respectively gq 6= 0. Thus
V =
p
M
g+j .
j=q
63
Let us now use a W -invariant inner product on E to identify E = E. Recall that under
2
. The transformation
this identification, = (,)
w () = 2
(, )
(, )
64
Lemma 15.14. An element x E is regular if and only if its stabilizer under the action of
W is trivial.
Proof. If x is not regular, there exists a root with (, x) = 0. It then follows that w (x) = x.
If x is regular, and w(x) = x, we will show that w = 1. Denote by h it be the element
corresponding to x under the identification E = Hom(u(1), t)
= it; thus d(h) = (, x) for all
X (T ). Since ker(ad(h)) gC is invariant under the adjoint representation of T G, it is
a sum of weight spaces. But ad(h) acts on the root space g as a non-zero scalar d(h) = (, x).
Thus ker(ad(h)) does not contain any of the root spaces, proving that ker(ad(h)) = g0 = tC .
It follows that t is the unique maximal abelian subalgebra containing ih. Equivalently, the
1-parameter subgroup S generated by the element ih t is contained in a unique maximal
torus, given by T itself. Suppose now that w W with wx = x, and let g N (T ) be a lift
of w. Then Adg (h) = h, so that g ZG (S). By our discussion of maximal tori, there exists a
maximal torus T containing S {g}. But we have seen that T is the unique maximal torus
containing S. Hence g T = T , proving that w = 1.
Remark 15.15. This result (or rather its proof) also has the following consequence. Let greg g
be the set of Lie algebra elements whose stabilizer group
G = {g G| Adg () = }
under the adjoint action is a maximal torus, and gsing = g\greg those elements whose stabilizer
is strictly larger than a maximal torus. Then
greg t
is the set of all t such that d() 6= 0 for all roots .
Exercise 15.16. For arbitrary t, the stabilizer G = {g G| Adg () = } contains T , hence
gC
is a sum of weight spaces. Which roots of G are also roots of G ? What can you say about
the dimension of G ?
The connected components of the set E reg are called the open Weyl chambers, their closures
are called the closed Weyl chambers. Unless specified differently, we will take Weyl chamber to
mean closed Weyl chamber. Note that the Weyl chambers C are closed convex cones. (That is,
if x, y C then rx + sy C for all r, s 0.) The Weyl group permutes the set of roots, hence
it acts by permutation on the set of root hyperplanes H and on the set of Weyl chambers.
Lemma 15.17. The Weyl group acts freely on the set of Weyl chambers. That is, if C is a
chamber and w W with wC C then w = 1.
Proof. If wC = C, then w preserves the interior of C. Let x int(C). Then wi x int(C) for
all i 0. Letting k be the order of w, the element x := x + wx + . . . wk1 x int(C) satisfies
wx = x . By the previous Lemma this means w = 1.
Exercise 15.18. Let C be a fixed (closed) Weyl chamber. a) Let D C one of its faces. (Thus
D is the intersection of C with some of the root hyperplanes.). Show that if w(D) D, then
wx = x for all x D. (Hint: D can be interpreted as the Weyl chamber of a subgroup of G.)
b) Show that if w W takes x C to x C then x = x.
65
66
i ki i
= 0 for some
kj <0
Taking the scalar product with itself, and using (i , j ) 0 for i 6= j we obtain
X
0 (, ) =
ki kj (i , j ) 0.
ki >0, kj <0
P
Hence
P = 0. Taking the inner product with any x int(C+ ), we get 0 = ki >0 ki (i , x ) =
kj <0 kj (j , x ). Hence the set of i with ki > 0 is empty, and so is the set of i with ki < 0.
Thus all ki = 0, proving that the i are linearly independent.
P
We claim that any R+ can be written in the form =
ki i for some ki Z0 .
This will prove (b), and also finish the proof of (a). Suppose the claim is false, and let be
a counterexample with (, x ) as small as possible. Since is not a simple root, it can be
written as a sum = + of two positive roots. Then ( , x ), ( , x ) are both strictly
positive, hence they are strictly smaller than their sum (, x ). Hence, neither nor is a
counterexample, and each can be written as a linear combination of i s with coefficients in
Z0 . Hence the same is true of , hence is not a counterexample. Contradiction.
spanZ R X (T ).
2 1
1 2
0 1
A=
0
0
1
0
0
0
2 1
1 2
This is also the Cartan matrix for SU(n) (which has the same roots as U(n)).
67
Example 16.6. Let G = SO(2l + 1). Using the standard maximal torus and the basis X (T ) =
spanZ (1 , . . . , l ), we had found that the roots are i j for i 6= j, together with the set of
all i . Let x = n1 + (n 1)2 + . . . + n . Then (x , ) 6= 0 for all roots . The positive
roots are the set of all i j with i < j, together with all i + j for i 6= j, together with all
i . The simple roots are
= {1 2 , 2 3 , . . . , l1 l , l }.
Here is the Cartan matrix for l = 4:
2 1 0
0
1 2 1 0
A=
0 1 2 1
0
0 2 2
= {1 2 , 2 3 , . . . , l1 l , 2l },
with Cartan matrix the transpose of that of SO(2l + 1).
A more efficient way of recording the information of a Cartan matrix is the Dynkin diagram 8 .
The Dynkin diagram is a graph, with vertices (nodes) the simple roots, connected by edges
if i 6= j and (i , j ) 6= 0. One gives each edge a multiplicity of one, two, or three according to
||i ||2
whether ||
2 equals 1, 2 or 3. For edges with multiplicity 2 or 3, one also puts an arrow from
j ||
longer roots down to shorter roots. Note that the Dynkin diagram contains the full information
of the Cartan matrix.
Example 16.7. There are only four possible Dynkin diagrams with 2 nodes: a disconnected
Dynkin diagram (corresponding to SU(2) SU(2) or SO(4)), and three connected ones, for the
three possible multiplicities of the edge connecting the two nodes. The Dynkin diagram for
multiplicity 1 is that of SU(3), the one for multiplicity 2 is that of SO(5). It turns out that
there is a unique compact, simple Lie group corresponding to the Dynkin diagram with two
nodes and an edge with multiplicity 3: This is the exceptional Lie group G2 .
Exercise 16.8. Using only the information from the Dynkin diagram for G2 , give a picture of
the root system for G2 . Use the root system to read off the dimension of G2 and the order of
its Weyl group. Show that the dual root system R for G2 is isomorphic to R.
Proposition 16.9. The positive Weyl chamber can be described in terms of the simple roots
as
C+ = {x E| (i , x) 0, i = 1, . . . , l}.
Proof. By definition, C+ is the set of all x with (, x) 0 for R+ . But since every positive
root is a linear combination of simple roots with non-negative coefficients, it suffices to require
the inequalities for the simple roots.
8Dynkin diagrams were used by E. Dynkin in his 1946 papers. Similar diagrams had previously been used
68
R = W .
Hence, given the Dynkin diagram one may recover the root system, the Weyl group, the
Weyl chamber etc.
Example 16.11. The Dynkin diagram of SO(5) has two vertices 1 , 2 , connected by an edge
of multiplicity 2 directed from 1 to 2 . Thus ||1 ||2 = 2||2 ||2 , and the angle between 1 , 2
2
is 3
4 . It is standard to work with a normalization where the long roots satisfy |||| = 2.
2
1
2
2
A concrete realization as a root system in R is given by 1 = and 2 = ; other
realizations are related by an orthogonal transformation of R2 .
2
1
2
s2 (l1 1 + l2 2 ) = l1 1 l2 2 ,
Hence
s1 (1 ) = 1 = 1 + 2 ,
s1 (2 ) = 1 ,
s2 (1 ) = 1 + 2
s2 (2 ) = 2 ,
s2 s1 (1 ) = 1 2 ,
s1 s2 (2 ) = 1 ,
which recovers all the roots. The Weyl group is the reflection group generated by s1 , s2 . As an
abstract group, it is the group generated by s1 , s2 with the single relation (s1 s2 )3 = 1.
69
For any root = sumli=1 ki i R (or more generally for any element of the root lattice),
one defines its height by
l
X
ki .
ht() =
i=1
l
X
hi , i.
ht() =
i=1
Since all ki 0, there must be at least one index r with (, r ) > 0. This then implies that
r R. Since 6 , there must be at least one index i 6= r with ki > 0. Since the
coefficient of this i in r is again ki > 0, it follows that r R+ .
17. Serre relations
Let G be a compact connected semi-simple Lie group, with given choice of maximal torus
T and positive Weyl chamber C+ . Let = {1 , . . . , l } be the set of simple roots, and
aij = h
i , j i be the entries of the Cartan matrix. Let hi [gi , gi ] with normalization
di (hi ) = 2. Pick ei gi , normalized up to U(1) by the condition [ei , ei ] = hi , and put
fi = ei .
Proposition 17.1. The elements ei , fi , hi generate gC . They satisfy the Serre relations,
(S1)
[hi , hj ] = 0,
(S2)
[ei , fj ] = ij hi ,
(S3)
[hi , ej ] = aij ej ,
(S4)
[hi , fj ] = aij fj ,
(S5)
(S6)
Proof. Induction on height shows that all root spaces g for positive roots are in the subalgebras
generated by the ei , fi , hi . Indeed, if R+ we saw that = + r for some R+ with
ht() = ht() = 1, and [er , g ] = g+r since r , , are all roots). Similarly the root spaces
for the negative roots are contained in this subalgebra, and since the hi span tC , it follows that
the subalgebra generated by the ei , fi , hi is indeed all of gC . Consider next the relations. (S1)
is obvious. (S2) holds true for i = j by our normalizations of ei , fi , hi , and for i 6= j because
[gi , gj ] gi j = 0 since i j is not a root. (S3) and (S4) follow since ej , fj are in the
root spaces gj :
[hi , ej ] = dj (hi )ej = h
i , j iej = aij ej
70
and similarly for [hi , fj ]. For (S5), consider the i -root string through j . Since j i is not
a root, the length of the root string is equal to k + 1 where k is the eigenvalue of ad(hi ) on
gj . But this eigenvalue is dj (hi ) = aij . Hence root string has length 1 aij , and consists of
the roots
j , j + i , . . . , j aij i .
In particular, j + (1 aij )i is not a root. This proves (S5), and (S6) is verified similarly.
The elements ei , fi , hi are called the Chevalley generators of the complex Lie algebra gC .
It turns out that the relations (S1)-(S6) are in fact a complete system of relations. This is
a consequence of Serres theorem, stated below. Hence, one may reconstruct gC from the
information given by the Dynkin diagram, or equivalently the Cartan matrix aij = h
i , j i.
In fact, we may start out with any abstract Root system.
We begin with the abstract notion of a root system.
Definition 17.2. Let E be a Euclidean vector space, and R E\{0}. For R define
= 2/(, ). Then R is called a (reduced) root system if
(a) spanR (R) = E.
(b) The reflection s : 7 h , i preserves R.
(c) For all , R, the number ( , ) Z,
(d) For all R, we have R R = {, }.
The Weyl group of a reduced root system is defined as the group generated by the reflections
s .
As in the case of root systems coming from compact Lie groups, one can define Weyl chambers, positive roots, simple roots, and a Cartan matrix and Dynkin diagram.
Theorem 17.3 (Serre). Let = {1 , . . . , l } be the set of simple roots of a reduced root system
of rank l, and let aij = h
i , j i be the Cartan matrix. The complex Lie algebra with generators
ei , fi , hi , i = 1, . . . , l and relations (S1)-(S6) is finite-dimensional and semi-simple. It carries
a conjugate linear involution 0 , given on generators by
0 (ei ) = fi , 0 (fi ) = ei , 0 (hi ) = hi ,
hence may be regarded as the complexification of a real semi-simple Lie algebra g. The Lie
algebra g integrates to a compact semi-simple Lie group G, with the prescribed root system.
For a proof of this result, see e.g. V. Kac Infinite-dimensional Lie algebras or A. Knapp,
Lie groups beyond an introduction.
18. Classification of Dynkin diagrams
There is an obvious notion of sum of root systems R1 E1 , R2 E2 , as the root system
R1 R2 in E1 E2 . A root system is irreducible if it is not a sum of two root systems.
Given an abstract root system, we may as before define Weyl chambers, and the same proof
as before shows that for non-orthogonal roots , with |||| ||||, the ratio of the root lengths
is given by ||||2 /||||2 {1, 2, 3}, and the angles in the three cases are 3 , 4 , 6 mod .
Hence, we may define simple roots and a Dynkin diagram as before.
Proposition 18.1. A root system is irreducible if and only if its Dynkin diagram is connected.
71
Proof. Let be a set of simple roots for R. If R is a sum of root systems R1 and R2 , then
1 = R1 and 2 = R2 are simple roots for Ri . Since all roots in 1 and orthogonal to
all roots in 2 , the Dynkin diagram is disconnected. Conversely, given a root system R E
with disconnected Dynkin diagram, then = 1 2 where all roots in 1 are orthogonal to
all roots in 2 . This gives an orthogonal decomposition E = E1 E2 where E1 , E2 is the space
spanned by roots in 1 , 2 . The simple reflections si for roots i 1 commute with those of
roots j 2 , hence the Weyl group is a direct product W = W1 W2 , and R is the sum of
R1 = W1 1 and R2 = W2 2 .
Hence, we will only consider connected Dynkin diagrams. The main theorem is as follows:
Theorem 18.2. Let R be an irreducible root system. Then the Dynkin diagram is given by
exactly one of the following diagrams
Al (l 1), Bl (l 2), Cl (l 3), Dl (l 4), E6 , E7 , E8 , F4 , G2 .
PICTURE
Here the subscript signifies the rank, i.e. the number of vertices of the Dynkin diagram.
We will sketch the proof in the case that the root system is simply laced, i.e. all roots have
the same length and hence the Dynkin diagram has no multiple edges. We will thus show that
all simply laced connected Dynkin diagrams are of one of the types Al , Dl , E6 , E7 , E8 .
We will use the following elementary Lemma:
Lemma 18.3. Let u1 , . . . , uk be pairwise orthogonal vectors in a Euclidean vector space E.
For all v E we have
k
X
(v, ui )2
2
,
||v|| >
||ui ||2
i=1
Proof in the simply laced case. We normalize the inner product on E so that all roots satisfy
||||2 = 2. Since all roots have equal length, the angle between non-orthogonal simple roots is
2
2
1
3 . Since cos( 3 ) = 2 , it follows that
(i , j ) = 1
if i , j are connected by an edge of the Dynkin diagram.
A subdiagram of a Dynkin diagram is obtained by taking a subset of vertices, together
with the edges connecting any two vertices in . It is clear that such a subdiagram is again a
Dynkin diagram. (If corresponds to the root system R, then corresponds to a root system
R spanR .)
The first observation is that the number of edges in the Dynkin diagram is < l. Indeed,
0 < ||
l
X
i=1
i ||2 = 2l + 2
(i , j ) = 2l 2#{edges}.
i<j
Hence #{edges} < l. Since this also applies to subdiagrams of the Dynkin diagram, it follows
in particular that the diagram cannot contain any loops.
72
One next observes that the number of edges originating at a vertex is at most 3. Otherwise,
there would be a star-shaped subdiagram with 5 vertices, with 1 , . . . , 4 connected to the central vertex . In particular, 1 , . . . , 4 are pairwise orthogonal. Since is linearly independent
of 1 , . . . , 4 , we have
2
2 = |||| >
4
X
(, i )2
4
X
1 2
(
) = 2,
=
2
4
X
(, i )2
i=1
||i ||2
i=1
a contradiction. (To get the inequality <, note that ||||2 is the sum of squares of its coefficients
in an orthonormal basis. The i /||i ||, i 4 is part of such a basis, but since is not in their
span we have the strict inequality.)
Next, one shows that the Dynkin diagram cannot contain more than one 3-valent vertex.
Otherwise it contains a subdiagram with a chain 1 , . . . , n , and two extra vertices 1 , 2
connected to 1P
and two extra vertices 3 , 4 connected to n . Let = 1 + . . . + n . Then
n1
2
|||| = 2n 2 i=1
(i , i+1 ) = 2, and (, i ) = 1. Hence, the same argument as in the
previous step (with here playing the role of 5 there) gives a contradiction:
2 = ||||2 >
i=1
||i ||2
4
X
1
( )2 = 2.
2
i=1
Thus, the only type of diagrams that remain are chains, i.e. diagrams of type Al , or star-shaped
diagrams with a central vertex and three branches of length r, s, t emanating from . Label
the vertices in these branches by 1 , . . . , r1 , 1 , . . . , s1 and 1 , . . . , t1 in such a way that
(1 , 2 ) 6= 0, . . . , (r1 , ) 6= 0 and similarly for the other branches. Let
=
r1
X
j=1
jj , =
s1
X
j=1
jj , =
t1
X
jj .
j=1
Then , , are pairwise orthogonal, and , , , are linearly independent. We have ||||2 =
r(r 1) and (, ) = (r 1), and similarly for , . Hence
2 = ||||2 >
(, )2
(, )2
(, )2
r1 s1 t1
+
+
.
+
+
=
||||2
||||2
||||2
r
s
t
Equivalently,
1 1 1
+ + > 1.
r s
t
One easily checks that the only solutions with r, s, t 2 and (with no loss of generality)
r s t are:
(2, 2, l 2), l 4, (2, 3, 3), (2, 3, 4), (2, 3, 5).
These are the Dynkin diagrams of type Dl , E6 , E7 , E8 . It remains to show that these Dynkin
diagrams correspond to root systems, but this can be done by explicit construction of the root
systems.
Consider the Dynkin diagram of E8 , with vertices of the long chain labeled as 1 , . . . , 7 ,
and with the vertex 5 connected to 8 . It may be realized as the following set of vectors in
R8 :
i = i i+1 , i = 1, . . . , 7
73
together with
8 = 12 (1 + . . . + 5 ) 21 (6 + 7 + 8 ).
(Indeed, this vectors have length squared equal to 2, and the correct angles.) The reflection si
for i 7 acts as transposition of indices i, i + 1. Hence S8 is embedded as a subgroup of the
Weyl group. Hence,
= 21 (1 + 2 + 3 ) + 12 (4 + . . . + 8 )
is also a root, obtained from 8 by permutation of 1, 2, 3 with 4, 5, 6. Applying s8 , we see that
s8 () = + 8 = 4 + 5
is a root. Hence, the set of roots contains all i j with i < j, and the Weyl group contains
all even numbers of sign changes. (In fact, we have just seen that the root system of E8 contains
that of D8 .) We conclude that
R = {i j } { 12 (1 2 8 )}
where the second set has all sign combinations with an odd number of minus signs. Note that
there are 2l(l 1) = 112 roots of the first type, and 27 = 128 roots of the second type. Hence
the dimension of the Lie group with this root system is 112 + 128 + 8 = 248. With a little extra
effort, one finds that the order of the Weyl group is |W | = 696, 729, 600.
19. Fundamental weights
Let G be a semi-simple compact Lie group. Inside E = X (T ) Z R, we have three lattices
spanZ (R ) X (T ) spanZ (R) ,
where spanZ (R) = Hom(spanZ (R), Z) is the dual of the root lattice.
Theorem 19.1. There are canonical isomorphism,
spanZ (R) /X (T )
= Z(G)
and
X (T )/ spanZ (R )
= 1 (G).
In particular, if G is simply connected the co-weight lattice agrees with the co-root lattice,
and hence has basis
i the simple co-roots. Furthermore, the center is then given in terms of
root data by
Z(G) = spanZ (R) / spanZ (R ).
We will give a full proof of the first part, but only an incomplete proof of the second part.
Partial proof. We have
Z(G) =
ker(),
since t T lies in Z(G) if and only if its action on all root spaces g is trivial. Under the
identification X (T ) Z R
= it, the elements of spanZ (R) consist of all it such that
d() Z for all R, hence exp(2i) Z(G). On the other hand, X (T ) corresponds to
elements such that exp(2i) = 1. This shows spanZ (R) /X (T ) = Z(G).
The inclusion T G gives a group morphism
X (T ) = 1 (T ) 1 (G).
74
Choose an invariant inner product B on g, and let G carry the corresponding Riemannian
e Given x 1 (G), let (t) be a geodesic in G
e from
metric. Think of 1 (G) as a subgroup of G.
0 to x. Its image (t) = expG (t) is then a closed geodesic in G representing the given element
of the fundamental class. But any geodesic of G is of the form exp(t) for some g. Choose
g G such that Adg () t. Then Adg (t) is homotopic to (t) (since G is connected), and
lies in T . This shows that the map 1 (T ) 1 (G) is surjective.
Recall next that we defined the co-roots as the composition of U(1)
= T SU(2)
with the map j : SU(2) G. But SU(2) is simply connected. Hence the loop in SU(2) is
contractible, and so is its image under j . This shows that spanZ (R ) lies in the kernel of the
map X (T ) 1 (G). The harder part is to show that it is actually equal to the kernel. For a
proof of this part, see e.g. Broecker-tom Dieck, chapter V.7.
Exercise 19.2. Find a generalization of this Theorem to arbitrary compact connected Lie
groups.
If G is semi-simple, define elements i E = X (T ) Z R by h
i , j i = ij . By definition,
these elements span the positive Weyl chamber C+ E. (If G simply connected, the Theorem
1
i
i < l 1,
+ . . . +
1
1
l1
l
i = i = 2 ( + . . . + ) i = l 1,
1 1
l1 + l ) i = l.
2 ( + . . . +
P i
P
The root lattice spanZ (R) consists of all
ki with integer coefficients such that
ki 2Z.
If l is odd, the quotient spanZ (R)/ spanZ (R) is isomorphic to Z4 ; it is generated by the image
of l . If l is even, we have spanZ (R)/ spanZ (R) = Z2 Z2 , generated by the images of l1
and l . Hence, the simply connected Lie group corresponding to Dl has center Z4 if l is odd,
and Z2 Z2 if l is even.
20. More on the Weyl group
We will need a few more facts about the relation of the Weyl group with the root system.
Recall that si = wi are the simple reflections corresponding to the simple roots. We have:
Lemma 20.1. The simple reflection si preserves the set R+ \{i }.
75
(6)
has coefficients kj = kj for j 6= i. Since is not a multiple of i it follows that kj = kj > 0 for
some j 6= i. This shows that si is positive.
Remark 20.2. This Lemma may also be understood geometrically. Fix int(C+ ), so that a
root is positive if and only if (, ) > 0. The Weyl group element si acts by reflection across the
hyperplane corresponding to i . Then si () lies in the adjacent chamber, and the line segment
from to si () does not meet any root hyperplane other than Hi = Hi . Hence, for all roots
6= i , the inner products (, ) and (, si ) = (si , ) have the same sign.
We have seen that the Weyl group is generated by simple reflections. We use this to define:
Definition 20.3. The length l(w) of a Weyl group element w W is the smallest number r such
that w can be written in the form
(7)
w = si1 . . . sir .
(The last identity follows e.g. since (1)l(w) is the determinant of w as an element of GL(E).)
It is also clear that for any i, l(wsi ) l(w) = 1. The sign can be determined as follows.
Proposition 20.4. For any Weyl group element w, and any simple root i , we have
(
l(w) + 1 wi R+
l(wsi ) =
l(w) 1 wi R
Proof. Consider a reduced expression (7) for w. Suppose wi R . Then there is an index
m r such that
sim+1 sir i R+ , sim sir i R .
In terms of u = sim+1 sir , these equations read
ui R+ , sim ui R .
Since im is the unique positive root that becomes negative under sim , it follows that im = ui .
Consequently,
sim = usi u1 .
Multiplying from the right by u, and from the left by si1 sim , we obtain
si1 sim1 sim+1 sir = wsi .
This shows l(wsi ) = l(w) 1. The case that wi R+ is reduced to the previous case, since
w i with w = wsi is negative.
76
For any w W , let R+,w be the set of positive roots that are made negative under w1 .
That is,
R+,w = R+ wR .
(8)
Proposition 20.5. For any w W ,
l(w) = |R+,w |.
(9)
(10)
77
In particular, the expression on the right hand side lies in the lattice spanZ (R ) spanned by
the fundamental weights (equal to X (T ) is G is simply connected). The element plays an
important role in representation theory.
21. Representation theory
Let G be a compact, connected Lie group. Fix a choice of maximal torus T and a positive
Weyl chamber C+ E = X (T ) Z R, and let R+ be the set of simple roots and positive
roots, respectively. Recall that the Weyl chamber is given in terms of simple co-roots by
C+ = { E| h , i 0, },
and that any E is W -conjugate to a unique element of C+ . In particular, every X (T )
is W -conjugate to a unique element in
X (T )+ := C+ X (T ).
We call these the dominant weights of G. If G is simply connected, so that X (T ) has basis the
P
fundamental weights, the dominant weights are those of the form li=1 mi i with mi Z0 .
That is, X (T )+ is a free monoid with basis the fundamental weights.
Similar to the representation theory of sl(2, C) (or equivalently SU(2)), the representation
theory of compact Lie groups relies on the concept of a highest weight. It is convenient to
introduce the nilpotent subalgebras
M
M
n+ =
g , n =
g .
R+
Then
gC = n tC n+
as vector spaces. In terms of the Chevalley generators ei , fi , hi , n+ is generated by the ei , n
is generated by the fi .
Recall that for any G-representation, we denote by (V ) the W -invariant set of its weights.
Theorem 21.1. Let : G End(V ) be an irreducible complex G-representation. Then there
is a unique weight (V ) X (T ), with the property that + is not a weight for all
positive roots R+ . The weight has multiplicity 1. All other weights (V ) are of the
form
X
ki i
=
i
where all ki Z0 .
78
P
Since fi takes V to Vi , each fi1 fir v is a weight vector, of weight li=1 ki i , where
P
ki = #{j : ij = i}. If r > 0, we have li=1 ki > 0. This shows that has multiplicity 1, and
that it is the unique highest weight of V .
79
The weight lattice X (T ) is obtained from the weight lattice X (T ) = spanZ (1 , . . . , l+1 ) by
P
j
projection along the direction l+1
j=1 .
i
i+1
The simple roots are i = , i = 1, . . . , l. For the standard inner product on E
these have length squared equal to 2, hence they coincide with the co-roots. The fundamental
weights dual to the co-roots are
l+1
i = (1 + + i )
i X j
.
l+1
j=1
The defining action of U(l + 1) on Cl+1 has the set of weights (V ) = {1 , . . . , l+1 } X (T ).
To get the corresponding set of weights (V ) of SU(l+1), we have to project along 1 +. . .+l+1 .
That is,
1
(1 + . . . + l+1 ), i = 1, . . . , l + 1}.
(Cl+1 ) = {i
l+1
Taking inner products with the simple roots, we see that only one of these weights lies in the
1
(1 + . . . + l+1 ) = 1 . We hence see that Cl+1
Weyl chamber C+ : This is the weight 1 l+1
is the irreducible representation of highest weight 1 .
Consider more generally the k-th exterior power of the defining representation, k (Cl+1 ).
The weights for the U(l + 1)-action are all i1 + . . . + ik such that i1 < < ik . Each of these
80
weights has multiplicity 1. The corresponding weights of SU(l + 1) are their projections:
k
(1 + . . . + l+1 ).
i1 + . . . + ik
l+1
Again, we find find that only one of the weights lies in C+ , this is the weight
k
1 + + k
(1 + . . . + l+1 ) = k .
l+1
That is, the irreducible representation of highest weight k is realized as k (Cl+1 ).
Proposition 21.7. The irreducible representation of SU(l + 1) of highest weight k is realized
as the k-th exterior power k (Cl+1 ) of the defining representation on Cl+1 .
It is also interesting to consider the symmetric powers of the defining representation.
Proposition 21.8. The k-th symmetric power S k (Cl+1 ) of the defining representation of
SU(l + 1) is irreducible, of highest weight k1 . Similarly the representation S k ((Cl+1 ) ) is
irreducible, of highest weight kl .
The proof is left as an exercise.
The complexified adjoint representation on the complexified Lie algebra su(l + 1)C is irreducible, since SU(l + 1) is simple. Its weights are, by definition, the roots together with 0.
Hence, there must be a unique root max such that + i is not a root, for any i. Indeed, one
observes that
max = 1 l+1
is such a root. It is the highest root, i.e. the root for which ht() takes on its maximum. In
fact, ht(max ) = l since it is the sum of the simple roots.
22. Highest weight representations
In this section we will present a proof of Weyls theorem for the case of a simply connected
compact Lie group G. The proof is on the level of Lie algebras. Hence, we will consider
complex representations : g End(V ) of the Lie algebra - if V is finite-dimensional, any
such representation integrates to the group level.
Let E = X (T ) Z R. It is convenient to identify E C
= (tC ) , by the linear map taking
C
X (T ) to the linear functional d : t C. Using this identification, we will simply write
in place of d, and likewise for the roots.
As before, we let n denote the nilpotent Lie subalgebra of gC given as the direct sum of
root spaces g for R . Their direct sum with tC is still a Lie subalgebra of gC , called the
Borel subalgebras b = tC n .
Definition 22.1. Let : gC End(V ) be a complex representation, possibly infinite-dimensional.
A non-zero vector v V is called a weight vector if span(v) is invariant under the action of tC .
For (tC ) we denote V = {v V | ()v = ()v, tC }. Then the weight vectors are
the non-zero elements of the V s. Note that need not lie in X (T ) in general.
Definition 22.2. A gC -representation V is called a highest weight representation if there is a
weight vector v with (n+ )v = 0 and such that
V = (U gC )v.
81
Thus, span(v) is invariant under the action of U (b+ ). Using the PBW isomorphism U (n )
U (b+ ) U (gC ) given by multiplication, it hence follows that
V = U (n ).v.
Thus, any highest weight representation is spanned by weight vectors vectors
fi1 fir .v
P
The weight of such a weight vector (11) is
ki i where ki = #{j| ij = i}, and the
multiplicity
of
any
weight
is
bounded
above
by
the
of sequences i1 , . . . , ir with =
Pr
.
As
a
very
rough
estimate
the
multiplicity
is
k as
Proposition 22.4. Let V be a highest weight representation, of highest weight (tC ) . Then
there exists a surjective g-module morphism L() V .
Proof. Let v V be a highest weight vector. The map C V, 7 v is b+ -equivariant.
Hence, the surjective g-map U (gC ) V x 7 (x)v descends to a surjective g-map L()
V.
Lemma 22.5. The Verma module L() for (tC ) has a unique maximal proper submodule
L (). The quotient module
V () = L()/L ()
is an irreducible highest weight module.
82
Proof. Any submodule is a sum of weight spaces; the submodule is proper if and only if does
not appear as a weight. Hence, the sum of two proper submodules of L() is again a proper
submodule. Taking the sum of all proper submodules, we obtain a maximal proper submodule
L ().
Now let V ( mu) = L()/L (). The preimage of a proper submodule W V () is a
proper submodule in L(), hence contained in L (). Thus W = 0. This shows that V () is
irreducible.
Proposition 22.6. Let V be an irreducible g-representation of highest weight (tC ) . Then
V is isomorphic to V (); the isomorphism is unique up to a non-zero scalar.
Proof. This is proved by a similar argument as in the Theorem for G-representations; see
Theorem 21.3.
It remains to investigate which of the irreducible gC -modules V () are finite-dimensional.
To this end we need:
Proposition 22.7. Let : sl(2, C) = spanC (e, f, h) End(V ) be a finite-dimensional representation. Then the transformation
= exp((e)) exp((f )) exp((e)) GL(V )
is well-defined. It implements the non-trivial Weyl group element, in the sense that
(12)
(h) 1 = (h),
1 1
0 1
1 0
1 1
1 1
0 1
0 1
1 0
83